Codebase knowledge graph for AI agents — 162 languages, sub-ms queries, 99% fewer tokens.
io.github.DeusData/codebase-memory-mcp MCP Server
This MCP server provides a codebase knowledge graph for AI agents. It is described as supporting 159 languages and sub-millisecond queries, with claims of “99% fewer tokens.” The repository centers on code analysis and code intelligence using knowledge-graph and AST-oriented tooling.
🛠️ Key Features
Codebase knowledge graph for AI agents
159 languages supported
Sub-ms query performance
“99% fewer tokens” claim
Includes tooling aligned with AST/tree-sitter and graph visualization
🚀 Use Cases
Code analysis and code intelligence for agents
Knowledge-graph queries over code
Graph visualization and codebase exploration
Cypher-based querying (as referenced by topics)
⚡ Developer Benefits
Developer tools oriented toward code intelligence
Model Context Protocol (MCP) server integration
Suitable for workflows involving code editors/agents (topics reference Claude Code, Cursor, Windsurf, Aider)
⚠️ Limitations
Supported capabilities are only described at a high level (knowledge graph, languages, query speed, token reduction) without detailed constraints in the provided excerpt.
Understand your repository, from source code to the systems around it. codebase-memory-mcp is an MCP server that connects code, documentation, configuration, and infrastructure in a persistent local knowledge graph. Use it to find an unfamiliar feature, follow a call across services, trace a deployment back to its source, or assess which parts of a project a change could affect.
Give your coding agent structural context through Model Context Protocol (MCP), or query the graph directly from the CLI. The MCP server is the core of the project. Its optional visual companion, CBM Lens, is in open beta (v0.0.1), with a source reader, architecture maps, the Galaxy graph, coverage diagnostics, architecture decisions, and local code chat.
The native C backend exposes 17 MCP tools, parses languages and file formats with vendored grammars, and adds Hybrid LSP type resolution for supported languages. The installer supports 46 automatic or conditional client integrations, including client configuration, skills, agent profiles, and applicable hooks. Indexing and graph queries need no hosted service or API key. Focused graph queries can reduce repeated file reads and token use; see the measured results.
Local processing: Source analysis and graph storage run on your machine. CBM Lens can optionally download a model for browser AI after you enable it; inference stays on the device. The installer writes agent configuration for the clients you select. See Security for release verification and SECURITY.md for reporting issues.
See the graph with CBM Lens (open beta v0.0.1). Folders become nested platforms, files and symbols become selectable blocks, and arrows show indexed relationships. The recording opens src/foundation in the map of this repository, selects arena.c and opens its source. Height can show source size; hotspot gravity reflects static fan-in. Still image.
Galaxy. Explore the code knowledge graph in 3D. The recording follows the outgoing relationships of cbm_arena_alloc one layer further, switches to the hierarchy and shows the two-hop call path to cbm_alloc. Edge pulses show relationship direction; they do not represent live execution. Still: two dependency layers around cbm_max_file_bytes.
Code knowledge graph in CBM Lens Beta Galaxy: expand cbm_arena_alloc, switch to the hierarchy and follow its call path through arena_grow to cbm_alloc
Explore. Browse files in an IDE-style workspace, read syntax-highlighted source in Monaco, and keep the code graph alongside it. File impact shows what depends on the open file, and Code details shows the callers and calls of the symbol under the caret. Select code to ask the local chat about it.
More views: system structure and behavior
System structure groups symbols into components and lets you inspect their dependencies.
Behavior shows possible calls from a selected operation and the source evidence behind them.
The high-resolution GIFs show CBM Lens Beta on 20 October 2026; the still images were captured on 12–13 October. All show public source from this repository. See capture details for each capture's build, indexed scope and recording setup.
Why codebase-memory-mcp
Find the right code with context: combine symbol, text, and semantic search with callers, dependencies, source snippets, and file outlines. Targeted queries and compact, paged results can reduce token use and repeated file reads; see Performance for scoped measurements.
Connect documentation to implementation: index document sections, PDF text pages, architecture decisions, and supported references into code. Find the explanation alongside the symbol.
Follow the system beyond source files: trace supported relationships through Docker, Kubernetes, Helm, Kustomize, Terraform/OpenTofu, Terragrunt, Ansible, scripts, Make, and CI configuration.
Understand a change before running it: inspect affected symbols, compare indexed snapshots, and find related tests. Experimental Smart CI combines static relationships and per-test coverage for supported C/C++ test harnesses, falling back to broader runs when evidence is insufficient.
Explore visually with CBM Lens Beta: read source, follow call paths, inspect architecture and coverage gaps, manage architecture decisions, and ask a local model about selected code.
Equip your coding agents: install sets up supported client configuration, durable guidance, skills, Scout/Verify/Auditor profiles, and applicable context hooks. Support varies by client; see the integration matrix.
Keep context local and reusable: indexes persist between sessions, eligible active MCP sessions can keep them synchronized, and optional graph artifacts let a team share a baseline.
Run a native tool on your machine: standard releases include CBM Lens for macOS, Linux, and Windows. The native runtime needs no Docker or language runtime; parser grammars are included.
The standard native release includes CBM Lens. Start it after installation with codebase-memory-mcp daemon start --open --port=9749. The Nix packages and source builds also offer builds without the UI.
Windows (PowerShell):
powershell
# 1. Download the installer
Invoke-WebRequest -Uri https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.ps1 -OutFile install.ps1
# 2. (Optional but recommended) Inspect the script
notepad install.ps1
# 3. Unblock the downloaded file (removes Mark-of-the-Web restriction added by browsers/Invoke-WebRequest)
Unblock-File .\install.ps1
# 4. Run it
.\install.ps1
Note: If you see a script execution policy error, run Set-ExecutionPolicy -Scope Process Bypass first, or invoke with PowerShell -ExecutionPolicy Bypass -File .\install.ps1.
Options: --skip-config (binary only, no agent setup), --dir=<path> (custom location).
Antivirus note: Microsoft Defender may flag a release binary as
Trojan:Script/Wacatac.B!ml. This is a known false positive; typically 61 of
~62 engines return clean, and the same detection family hits gh, llama.cpp,
Godot and Microsoft's own Go toolchain. See
Antivirus False Positives for the
evidence, how to verify the artifacts yourself, and how to report it if you
think we are wrong.
Restart your coding agent. Say "Index this project": done.
Manual install
Download the archive for your platform from the latest release:
codebase-memory-mcp-<os>-<arch>.tar.gz (macOS/Linux) or .zip (Windows)
Extract and install (each archive includes install.sh or install.ps1):
The install command automatically strips macOS quarantine attributes and ad-hoc signs the binary; no manual xattr/codesign needed.
Agent setup: skills, profiles, and hooks
Installation also equips supported coding agents to use the graph in daily work. The native install command detects clients and installs the integration pieces their documented formats support:
Component
What it adds
MCP configuration
Connects the client to the local server and its tools.
Instructions and skills
Explains graph-first exploration, index freshness, coverage checks, and when to read source.
Scout, Verify, and Auditor profiles
Offers fast provisional lookup, task-directed verification, or bounded full verification, where custom agents are supported.
Session and compaction hooks
Restores graph context at supported lifecycle events.
Search and read hooks
Adds relevant graph context to supported searches and flags recorded coverage gaps after supported file reads.
Hook augmentation is context-only and fail-open: it does not replace or block the requested tool call. Features vary by client and platform; some clients use durable instructions or a parent-agent handoff instead of hooks or direct subagent MCP access. Codex users review and trust installed hooks through /hooks. See Multi-Agent Support and Sessions, compaction, and subagents for exact support and activation requirements.
Hooks are enabled by default. Run codebase-memory-mcp install --no-hooks to remove CBM-managed hooks while keeping MCP registration and other integration files; the choice persists for later installs and updates. Run codebase-memory-mcp install --hooks to re-enable them, or inspect the current preference with codebase-memory-mcp config get install_hooks.
Session Coordination Daemon
CBM automatically shares one per-account coordination daemon across Claude Code, Codex, OpenCode, and every other configured client. There is no opt-in setting for MCP servers or hook clients: the first daemon-backed CBM session starts it, each session registers its own work, and a session-managed daemon stops after the final session. A daemon started with daemon start remains running until stopped. The daemon owns long-lived background services such as watchers, shared indexing jobs, and the optional UI. Closing one session cancels work owned only by that session, while work still needed by another session continues.
The detached daemon does not depend on an MCP frontend's stderr. It keeps owner-only durable records under the canonical ${CBM_CACHE_DIR}/logs directory (default ~/.cache/codebase-memory-mcp/logs):
File
Contents
cbm-daemon.log
Daemon lifecycle, watcher/indexing, UI, resource, and error events.
daemon-conflicts.ndjson
Exact-build, coordination-ABI, and cache-root admission conflicts.
activation-events.ndjson
Install/update/uninstall activation progress and outcomes.
Thin frontends still write immediate startup and session-specific errors to their own stderr; MCP JSON-RPC stdout remains clean.
All active CBM processes must run the exact same version, executable build, coordination ABI, and canonical cache root. Equivalent CBM_CACHE_DIR aliases resolve to the same root; a genuinely different root is rejected while any CBM process is active. MCP servers, hooks, one-shot CLI commands, temporary index workers, and the daemon share a crash-safe OS admission barrier; starting an ordinary conflicting process fails before doing work and records an explicit conflict in ${CBM_CACHE_DIR}/logs/daemon-conflicts.ndjson.
The native install and uninstall commands are the deliberate exception to that conflict rule. For installation, download, verification, and private same-filesystem staging finish before activation so a bad candidate never disrupts active work. The native update command prints the installer command that performs the update. Activation then publishes account-wide maintenance intent, asks the daemon and every temporary local operation to cancel, and waits to a finite deadline for all coordinated CBM processes to exit. It holds the admission and lifetime barriers exclusively while changing the active binary, configuration, PATH, or indexes. New CBM work cannot enter during this window. Activation progress and results are recorded in ${CBM_CACHE_DIR}/logs/activation-events.ndjson, and a successful command tells you to restart open coding-agent sessions so they launch the activated build.
Package-manager setup (npm, PyPI, or Go) verifies and caches the native executable for that package. It does not replace the active native installation and therefore does not stop running CBM sessions. When that cached binary is executed, it still enters the same exact-build admission barrier. The shell and PowerShell installers invoke the verified candidate's native install command, so they do receive the full account-wide activation guarantee.
The ordinary cli mode runs one command locally, but it is not disconnected from the daemon above: it connects to the shared coordination daemon (starting one if none is already running) for that same admission barrier plus per-project locks on graph mutations. Its connection is a cli_session: it does not register a project watcher or open a browser, and it holds the admission lease only for the command's lifetime. The shared daemon may serve Lens according to its UI configuration and available assets. If the command's own connection is what started the daemon, that daemon exits again once the command closes and no other session is attached; if a daemon was already running, the command simply joins and leaves it exactly as found. While the command is running, a temporary monitor separately lets activation cancel that operation and its supervised worker safely; the monitor exits with the command and never becomes a standing daemon. See CLI Mode for details.
CBM Lens Beta: code visualization and source navigation
CBM Lens is an optional visual companion to the MCP server, currently in open beta (v0.0.1). Standard native releases include its browser interface, so you can launch Lens directly after installation; UI-enabled source builds include it too. It builds on CodeAtlasWeb by Bernhard Jackiewicz (design and tracking in #1964). Lens is served on loopback. Browser AI downloads its optional model only after you enable it; inference stays on-device. Each panel names where its facts come from, and what the index did not record is shown as a gap. Start the local server:
The browser opens at http://127.0.0.1:9749. A permanent daemon keeps the UI available after the command exits. Use daemon status to inspect it and daemon stop when you are finished. For an MCP session, --ui=true --port=9749 enables the same shared UI. Active clients must use the same executable build and cache directory.
Open your first repository
Open the project menu and choose Add project index. You can also use Alt+P.
Choose a local repository folder and name the index. Files stay in place.
Start indexing and follow its reported status. You can close the dialog while it runs and reopen its activity from the project menu.
Open the indexed project. Start with Architecture for an overview or Explore to read a file.
To bookmark a view, use ?project=<name>&workspace=architecture. Other workspace values are explore, galaxy, adr, coverage, and system.
Workspace
What you can do
Explore
Navigate the indexed file tree and read source in Monaco. File impact shows what depends on the open file; Code details shows the callers, calls, references and related tests of the symbol under the caret.
Galaxy
Search and focus graph nodes, load their neighbors layer by layer, switch to hierarchy, follow a path, and open source. Back and Forward return to earlier views. The coverage shadow marks known gaps around the loaded graph.
Architecture
Explore the folder map, entry points, routes, hotspots, system structure, and behavior paths. Select a node or connection to inspect its evidence.
ADR
Read and edit the project's architecture decision record. Save explicitly; the document is stored with the index and shared with manage_adr.
Coverage
Check indexed, partially parsed, skipped, and excluded paths, together with available freshness information.
System
Inspect daemon activity, indexes, and logs. Filter logs by project, severity, or text.
Architecture views
Overview: move through source folders and files on a 3D map. Choose source-size or uniform height, language or kind colors, and relationship filters. Switch to Entry points to follow a static call graph from a starting symbol.
Routes: switch between the service map and endpoint connections. Deployment metadata and indexed route evidence are shown separately; unresolved registrations remain visible.
Hotspots: inspect symbols using measured signals such as fan-in and complexity. Gravity wells help locate frequently referenced areas. They are not a code-quality score.
System structure: inspect component dependencies, expand groups, isolate cycles, and follow connections back to source evidence.
Behavior: choose a source and a reachable destination to inspect static call paths and their source statements. These paths show possible relationships, not a recorded execution order. Runtime observations require separately ingested traces.
Drag a 3D view to orbit, scroll to zoom, and select a node or edge to inspect it. Edge colors distinguish relationships, and pulses indicate their direction. Reduced-motion preferences pause the pulses. Large views show their limits and omitted counts. A missing edge can reflect an indexing or display limit, so check Coverage and the source before drawing conclusions.
Chat with your codebase locally
Open Chat with the panel button at the top right to ask about the code you are looking at. In Explore it uses the open file or the exact text you have selected; in Galaxy it uses the selected node and its loaded neighbors. You can also attach graph evidence through Add context.
Click the robot icon at the top right to open the agent configuration.
Choose a local model. The default Qwen2.5 Coder 0.5B downloads about 567 MB; Qwen3 0.6B, LFM2.5 1.2B and Qwen3.5 2B are larger options. All need WebGPU with shader-f16 support.
Click Download & load. Model files come from Hugging Face; inference runs in your browser, with no hosted chat API or API key. A model that is already in the browser cache loads with Load model (cached, no download).
Ask a question. The composer shows the file or selection that will accompany it. Select a smaller range when you want to focus the answer.
Questions about callers and callees, such as "Who calls cbm_pipeline_run?", are answered from the loaded graph without the model, and the answer says so. When you select a symbol, the chat lists its indexed facts and the lines of its code, and the model adds one sentence; that sentence is left out when it claims something the source does not show. Text written by the model is labelled as such.
Chat with your codebase locally in CBM Lens Beta: callers and callees of arena_grow answered directly from the indexed graph
The chat history is kept for 24 hours in this browser, per project. Each new message uses the source currently open in the reader; earlier source snapshots are not silently accumulated, and messages about another file or selection are not sent along. If a prompt is too long for the model, older messages are left out first; if it still does not fit, CBM Lens asks you to select less code instead of cutting the code. This chat explains the context you give it; it cannot edit files or run tools. Check its answers against the source. The small default model words things loosely, especially in languages other than English.
Search
Cmd+K or Ctrl+K opens the Galaxy search. It finds symbols, files and folders by name or path in the whole index, also nodes that are not loaded yet. For callers and callees, ask in the chat; the answer is listed from the graph.
Architecture decisions
The ADR workspace keeps an unsaved draft per project in this browser tab. It checks whether the stored document changed before saving, but the API does not provide an atomic document lock. The editor accepts up to 8,000 UTF-8 bytes. Saving changes the indexed decision document, not a file in your repository.
Experimental agent activity
The Agents workspace and activity overlays are hidden in normal builds. To include them in a source build:
Activity comes from explicitly configured tool hooks. The bundled adapter supports Claude Code PostToolUse; other clients need their own adapter. Recorded events go to the local daemon, and the reader must enable activity before the browser polls them. This does not automatically observe every agent. See the frontend setup guide.
For troubleshooting, open System logs. Frontend errors and failed requests are recorded locally. GET /api/ui-log reads frontend entries from the daemon's SQLite journal; ${CBM_CACHE_DIR}/logs/ui.log is also retained as a rotating JSONL export. See the frontend README for log scope, model setup, and development details.
Auto-Index
Enable automatic indexing on MCP session start:
bash
codebase-memory-mcp config set auto_index true
auto_index is off by default. When enabled, new projects are indexed automatically on first connection. The configurable file limit is config set auto_index_limit 50000. Previously indexed projects can be watched through auto_watch (on by default), subject to its global setting and session ownership.
Each initialized, long-lived MCP session can register its root project when it connects and any target from a successful index_repository call, including custom project names and roots in another worktree. While auto_watch is enabled, these watches belong to the registering sessions and remain while at least one owning session is connected; closing the last owner removes the watch. Under a daemon started with daemon start a watch outlives its last owning session: the project stays watched and is re-indexed in the background when it changes, until that daemon stops, the project is deleted, its directory disappears, or auto_watch is set to false. An async index registers its target only after success and while its requesting session remains connected. CLI indexing and failed MCP indexes do not register watches. Global auto_watch controls registration and removes all watches on the next refresh or registration request when set to false. index_status reports this in its watch object: watched, and either the poll cadence and last scan time or the reason it is not watched.
Watcher registration is controlled globally by auto_watch (default true). On the next refresh or registration request, true allows MCP sessions to register their root on connect and their successful index_repository targets; false removes all watches. CLI indexing and failed MCP indexes do not add watches.
To turn the watcher off entirely, set config set watcher_enabled false (default true): the background poll thread never starts and no project is registered, while auto_index and manual index_repository keep working. Unlike auto_watch, which takes effect on the next refresh or registration request, watcher_enabled is read once when the background daemon starts, so run codebase-memory-mcp daemon stop after changing it; reconnecting your MCP client alone will not restart the daemon. See docs/CONFIGURATION.md.
The watcher follows git projects only by default: a project indexed from a directory that is not a git repository is not watched and stays at its last index until you run index_repository again. To have such roots polled too, set config set watch_non_git true (default false, read when the daemon starts): they are then checked on the same cadence with a file-tree scan that uses the indexer's own ignore rules.
Linked git worktrees are indexed as their own projects by default. Set config set ignore_worktrees true to skip them on the automatic paths — useful when short-lived git worktree add checkouts would otherwise each leave behind a permanent index of the same repository. An explicit index_repository call on a worktree is then refused unless you pass index_worktree=true. See docs/CONFIGURATION.md.
Keeping Up to Date
Updates run from the install script on every platform, not from inside the running binary.codebase-memory-mcp update validates your flags and then prints the exact command to run:
bash
# macOS / Linux
bash "<install-dir>/install.sh"
powershell
# Windows
powershell -File "<install-dir>\install.ps1"
The install script is placed next to the binary at install time, so the printed path resolves beside the executable. It is idempotent, so re-running it is the update: it stops the daemon, retires the running binary, installs the new one, and cleans up.
Why it works this way. On Windows it is a hard requirement; a running executable cannot replace its own image, so the swap has to happen from a process that is not the binary being replaced. On macOS and Linux it is a deliberate choice: an in-process updater is structurally a downloader (fetch an archive, verify it, unpack it, mark a file executable, run it), and shipping that composite in every binary to serve a command most people run a handful of times is a poor trade. The release archives now carry no download URLs at all, and cbm makes no network request of its own accord: it does not check for new versions in the background, and nothing phones home. You find out about releases from the install script, your package manager, or GitHub.
If PowerShell refuses to run the script because the file came from the internet, Unblock-File it first.
Installed through npm or pip? Update with your package manager on every platform (npm install -g codebase-memory-mcp@latest / pip install -U codebase-memory-mcp).
Uninstall
bash
codebase-memory-mcp uninstall
Removes owned agent config entries, skills, hooks, instructions, and the installed binary. Existing graph indexes are listed and kept by default. -y/--yes, -n/--no, and noninteractive input keep them unless --delete-indexes is explicitly given. An interactive terminal without that flag is asked separately; the default answer is to keep them.
bash
codebase-memory-mcp uninstall -y --delete-indexes # also delete every project index
codebase-memory-mcp uninstall --dry-run --delete-indexes # preview; change nothing
--delete-indexes is explicit consent and takes precedence over --no; --dry-run always preserves the files. When indexes are kept, uninstall prints their cache directory (${CBM_CACHE_DIR:-~/.cache/codebase-memory-mcp}) and how to remove them.
The install script placed beside the binary is reported, not deleted: uninstall prints its path and the rm command for it. It is left alone on purpose: it may be your own copy, a symlink into a checkout, or managed by a package manager, and an uninstaller should not delete a file it cannot prove it owns.
Features
Graph & analysis
Architecture overview: get_architecture returns a compact summary by default. Request specific aspects for routes, hotspots, dependencies, clusters, or cycles. system_structure provides the projection used by CBM Lens and must be requested on its own.
Architecture Decision Records: manage_adr persists the project's decision record across sessions; indexing also extracts supported ADR metadata and supersession from documentation files
Louvain community detection: Discovers functional modules by clustering call edges
Change impact: detect_changes maps uncommitted changes to affected symbols with risk classification and can report test impact with scope: "tests"; compare_graphs compares indexed snapshots
Runtime observations: ingest_traces adds observed call relationships from separately collected traces alongside static analysis; static call paths alone do not show what ran
Call graph: Resolves function calls across files and packages (import-aware, type-inferred)
Dead code candidates: find functions with no indexed callers, excluding entry points. Check dynamic dispatch, coverage gaps, and framework conventions before removing code.
Cypher-like queries: MATCH (f:Function)-[:CALLS]->(g) WHERE f.name = 'main' RETURN g.name
Search
Semantic search (the semantic_query parameter of search_graph, not a separate tool): vector search across the entire graph, powered by bundled Nomic nomic-embed-code embeddings (40K tokens, 768d int8) compiled into the binary; no API key, no Ollama, no Docker. 11-signal combined scoring (TF-IDF, RRI, API/Type/Decorator signatures, AST profiles, data flow, Halstead-lite, MinHash, module proximity, graph diffusion).
BM25 full-text search via SQLite FTS5 with cbm_camel_split tokenizer (camelCase / snake_case aware)
Code search (search_code): graph-augmented grep over indexed files only
Documentation connected to code
Find the explanation behind a symbol and follow supported references back to its implementation. Documentation participates in the graph alongside source files:
Markdown, reStructuredText, and AsciiDoc become navigable document sections; PDFs with text layers become searchable text pages. PDF extraction does not perform OCR.
Document-to-code references and supported language doc-comment references create MENTIONS relationships when the target can be resolved.
Architecture decision records contribute metadata such as status, deciders, and supersession, so decisions can be explored with the code they describe.
Supported diagrams and OpenAPI descriptions contribute structured entities. Extraction and reference resolution differ by format; not every textual mention becomes a graph edge.
Use graph and text search to locate the relevant section, follow its recorded relationships, and open the source evidence. Check check_index_coverage before treating a missing result as proof that the repository has no such documentation or reference.
Infrastructure and deployment
Ask "What deploys this service?", "Which image runs it?", or "Where does this configuration reach code?" CBM connects supported infrastructure forms to the source they build, configure, and run:
Containers and clusters: Dockerfiles, Compose services, Kubernetes resources, Helm charts and releases, and Kustomize overlays and patches.
Provisioning and configuration: Terraform/OpenTofu resources and modules, tfvars overrides, Terragrunt dependencies, Ansible tasks and handlers, and supported configuration keys and environment-variable access.
Build and delivery: shell commands, Make targets, and supported CI steps that build images, invoke repository scripts, or deploy resources.
For resolvable inputs, a trace can follow Service → workload → image → Dockerfile → program, or follow deployment and configuration relationships. Use trace_path modes for topology, deployment, and run paths. These are static relationships from supported source forms; dynamic values and unresolved templates can leave gaps. See Edge types for the exact relationship rules and limitations.
Cross-service linking
HTTP route ↔ call-site matching with confidence scoring
gRPC, GraphQL, tRPC service detection with protobuf Route extraction
Channel detection (EMITS / LISTENS_ON) for Socket.IO, EventEmitter, and generic pub-sub patterns across 8 languages with constant resolution
Cross-repo intelligence
CROSS_* edges link nodes across multiple repos indexed under the same store
Cross-repo architecture summary combining services, routes, and dependencies across the indexed fleet
Run status: index_status (and the UI layout next to linked_projects) carries a cross_repo object, {"status":"never_run"} until index_repository(mode="cross-repo-intelligence") has run with that project as the source, then "ran" with last_run_at, outcome, targets, projects_scanned, total_cross_edges and skipped_projects. That is how you tell "never linked" apart from "linked, nothing matched". The status is stored in the project's own index next to its CROSS_* edges, so a reindex that rebuilds the index drops both at once and the status goes back to never_run.
Skipped stores: target_projects: ["*"] skips an index built before the #768 schema change rather than failing the whole run. The response lists each one as skipped_projects: [{"name", "reason": "pre_768_schema", "hint": "reindex this project"}].
Smart CI and test impact
Use the graph to explain which tests a change reaches, and use that evidence to focus supported CI runs. Smart CI is experimental and has two layers:
Layer
What it does
What you supply
CI job selection
Chooses jobs from the changed-file list.
Repository-specific path and job rules; this repository's selector is an example.
Test selection
Combines static reachability, changed tests, and admitted per-test coverage to select suites or tests.
A supported C/C++ test harness, configuration, a coverage map, and a runner that accepts named selections.
codebase-memory-mcp test-impact select produces a selection with reasons; test-impact publish produces the graph-and-coverage bundle used as its baseline. For agent-assisted investigation, detect_changes with scope: "tests" exposes test impact through MCP.
Narrowing currently supports C/C++ family code with this repository's test conventions or declared equivalents. It is not general test selection for every language CBM indexes. Missing, stale, incompatible, or incomplete evidence broadens the selection or returns run_all; static reachability without admitted coverage never omits a suite. Begin with report-only shadow mode and validate your harness before enabling gating.
In this repository, test narrowing is on for every PR unless TEST_IMPACT_MODE says otherwise: the Linux x86_64 test leg runs the selection, and the arm64, macOS and Windows legs run it with every platform-specific suite whole; release and dry-run workflows retain full execution. See the Smart CI guide for supported conventions, coverage admission, setup, version requirements, and fallback behavior.
Edge types (selected)
CALLS: a callable is invoked at the source site
CALL_REFERENCE: a callable is used at a supported reference site (for example, a direct value argument) and resolves to one exact target
USAGE: an identifier is used, but a unique callable target is not proven (including ambiguous or complex expressions)
IMPORTS, DEFINES, IMPLEMENTS, INHERITS
HTTP_CALLS, ASYNC_CALLS (cross-service); a shell script, Make recipe or CI step running curl / wget on a URL it states (constant words resolved, unknown parts {}) to its Route with the method its flags give: a Kubernetes Service of the repository by the path, as code clients do, any other host (a download, localhost) with the host kept; the edge keeps via, method, url_path, host, never the user info or the query
SPAWNS: the caller starts another program (subprocess.run, exec.Command, posix_spawn, Process.Start, a shell command or Make recipe running a program by path, ...); the target is a Process node named after the program, or <dynamic> when the program is not a literal. When a shell script, Make recipe, GitHub Actions step or Prow job runs a script of the repository (./hack/x.sh, bash hack/x.sh), the target is that script's Module (else its File); make TARGET (also $(MAKE), -C, -f) targets the rule of that target, followed through included makefiles; edges from CI carry ci and job. A Dockerfile's last stage ENTRYPOINT (else CMD) spawns what the container runs (api: dockerfile): the repository script or source copied into the image, the Go main of a binary a stage builds with go build -o, else a Process
INFRA_MAPS: a Kubernetes Service to each workload whose pod labels carry every key: value of the Service's selector, within one namespace
CONNECTS_TO: a Kubernetes workload or compose service to the Service its containers' environment addresses (PRODUCT_CATALOG_SERVICE_ADDR: productcatalogservice:3550, a URL's host, <svc>.<ns>.svc..., a bare name on an *_ADDR / *_HOST / *_URL / ... key; compose: the project's .env interpolated): a Kubernetes Service of the same bundle and namespace, or a compose service of the same file. The edge keeps via: env, key and container, never the value
USAGE with ref: image: a Kubernetes workload (any pod template: containers, initContainers, CronJobs, custom resources) or a docker compose service to the Image it runs
Helm templates are rendered for their images: with the chart's values (the ChartDEPLOYS them, default_of: the values keys read) and with each release's (a Flux HelmRelease, a kustomize helmCharts entry, a Terraform helm_release: override_of). An image the render cannot know is left out, never guessed: after a change of values it does not run exactly, a tpl or include it cannot look into, a release value it cannot read, or a work budget it runs past. Known limit: each template file renders on its own, while Helm renders a chart's files with one shared .Values (a subchart's is its parent's .Values.<alias>), so a chart that changes .Values in one file and reads it in another can still show a wrong image
DEPLOYS: what puts manifests, kustomizations or charts on a cluster, to what it puts there: a Flux Kustomization or Argo Application to the kustomizations / manifests / chart of its path, a Flux HelmRelease to its chart (spec.chart or chartRef → OCIRepository), a shell command or Make recipe (kubectl apply -f/-k, kustomize build | kubectl apply, helm install/upgrade/template), a Go installer loading a chart with the Helm SDK, a Terraform helm_release, a Terraform kubernetes_* resource the object it declares (a Resource like a manifest's, ecosystem: terraform, when its name is a literal), a kustomization inflating a chart (helmCharts; its values file an IMPORTS). A chart outside the repository is an external Chart, keyed by its publisher when its repository URL is stated (__chart__<repository>/<name>, the same node in every project, the URL in repository), else one per project and name; the edge says via (flux, argo, helmrelease, kubectl, kustomize, helm, helm-sdk, terraform, helmcharts, ...)
DEPENDS_ON: a Helm chart to the charts it declares (the local chart for a resolvable file:// repository, otherwise an external Chart; the edge keeps repository), and a go.mod / requirements.txt file to an external Package; a Dockerfile to the Images it builds on (via: from, copy-from); an Image to the Dockerfile it is built from (via: build; tool: skaffold, compose, docker-build for docker / podman / nerdctl / buildah build commands in scripts, Make and CI); a compose service to the services its depends_on / links name
CONFIGURES: code reading an environment variable to the shared EnvVar of that name (strategy: env_access); a Kubernetes container, compose service or Dockerfile (ENV of the image) setting one to that container's own EnvVar (strategy: env_set, with source: value, configMapKeyRef, secretKeyRef, fieldRef, ..., compose, dockerfile; values are never stored); and that container's EnvVar to the code reading the name inside the files its image is built from (strategy: env_read, with dockerfile): the image's files are the sources its Dockerfile's COPY / ADD put into it, less the context's .dockerignore
CONFIGURES with strategy: key_symbol: a code symbol to the config key it names (config_key). A YAML, JSON, HCL (Terraform, OpenTofu, tfvars), TOML or INI key links by exact name (maxRetryCount = max_retry_count) and only to code inside the config file's component: its directory, or the one around a config/, configs/, conf/, settings/ or etc/ directory. A .env or .properties key also links by substring (confidence 0.75)
OVERRIDE with via: patch: a Kustomize patch to each object of its kustomization's build it changes (child → base): a patch file's documents to the object of the same kind and name; an inline patch (from the kustomization) or a JSON6902 operation file to what its target selects (group, version, kind, name, namespace, label selector); a component's in the build of each kustomization including it. The edge keeps kustomization, patch and line; a patch file is an IMPORTS (local_name: patches) of its kustomization, never a member of the build
OVERRIDE with via: namePrefix / nameSuffix: an object a kustomization renames, from its variant: a Resource named with the deployed name (capi- + controller-manager; the prefixes of nested kustomizations outermost first, suffixes innermost first) in the namespace that build gives it, ecosystem: kustomize, which the kustomization's file DEFINES. CRDs, Namespaces and API services keep their names, as kustomize does. The edge keeps kustomization and line
USAGE with via: terraform: a Terraform / OpenTofu reference (var.X, local.X, module.X, data.T.N, T.N) from the top-level block holding it to the block of its module (the .tf / .tofu files of its directory) that declares it (a locals entry: its locals block); an entry of the block's depends_on list is a DEPENDS_ON instead. The edge keeps ref (var, local, module, data, resource) and line; a name only mentioned in string text is no reference. A module call with a local source (module "net" { source = "./modules/net" }) IMPORTS (local_name: the module's name) the Folder of that directory (the root directory: the Project), and module.net.id (also module.net[0].id, module.net[*].id) USAGEs (ref: output) the output "id" there, beside module.net
OVERRIDE with via: tfvars: a .tfvars assignment to a variable the module of its own directory declares, from the file's Variable to that variable block (the edge keeps line); a .tfvars file whose own directory declares none of its names sets the nearest ancestor module's that does, the edge marked inferred (such a file is applied by a -var-file the repository does not state)
Terragrunt (via: terragrunt): a unit (terragrunt.hcl), evaluated in its own context: an includeIMPORTS (local_name) the file its path names; a read_terragrunt_configIMPORTS (read: the file holding the call) the file it names, from the block holding the call or, for a call in an included file, from the unit's include block that brings it in; a dependencyDEPENDS_ON the other unit; a local terraformsourceIMPORTS its Folder. Paths may use find_in_parent_folders, dirname and get_terragrunt_dir; anything else is unknown and links nothing
Ansible (via: ansible): a playbook or task file (a YAML file whose top level is a sequence) IMPORTS (local_name: the directive) what its directives name: import_playbook a playbook, include_tasks / import_tasks / include a task file (from its directory, else its role's tasks/), roles: items and include_role / import_role a role's Folder (roles/<name> at or above its directory). A templated target links nothing. Each task and handler is a Function (named by its name:, else its module; source: ansible, kind, module) that its file DEFINES: in a playbook's plays, a role's tasks/ and handlers/ files, and the files those include; a block's tasks, not the block. A notifyCALLS (via: notify) each handler of that name or listen topic in the task's own role (else its file). A kube (kubespray's module) or kubernetes.core.k8s task DEPLOYS (via: ansible, module) the repository file its play rendered the applied manifest from: the template (in the role's templates/) or copy (its files/) task whose dest is the task's filename, item by item through register / with_items / loop lists the play states (a literal list, a set_fact, the role's defaults / vars), or a k8stemplate: argument
EMITS, LISTENS_ON (channels)
DATA_FLOWS with arg-to-param mapping + field access chains
164 vendored tree-sitter grammars compiled into the binary
Generic package / module resolution: bare specifiers like @myorg/pkg, github.com/foo/bar, use my_crate::foo resolved via manifest scanning (package.json, go.mod, Cargo.toml, pyproject.toml, composer.json, pubspec.yaml, pom.xml, build.gradle, mix.exs, *.gemspec)
Infrastructure-as-code indexing: Dockerfiles (Dockerfile.*, *.Dockerfile, Containerfile: bases, what the container runs, ENV), container images (Image: the normalised repository, tag and digest on the edges; a repository built from several Dockerfiles, one per component tag, gets an Image per component), docker compose services (Resource), Kubernetes manifests (Resource, with group, version, kind, namespace, labels as k=v,k2=v2), Kustomize overlays (IMPORTS to their entries and patch files, OVERRIDE from their patches), Helm charts (Chart, DEPENDS_ON), Terraform / OpenTofu (.tf, .tofu, .tfvars) as graph nodes. A Service walks to the code it runs: Service → workload → Image → Dockerfile → main
Hybrid LSP semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl; a lightweight C implementation of language type-resolution algorithms, structurally inspired by and compatible with major language servers including tsserver / typescript-go, pyright, gopls, Roslyn, Eclipse JDT, and rust-analyzer (parameter binding, return-type inference, generic substitution, JSX component dispatch, JSDoc inference for plain JS files, namespace + trait + late-static-binding resolution for PHP, file-scoped namespaces + records + LINQ method syntax for C#, class-hierarchy + overload + lambda resolution for Java, extension-function + scope-function resolution for Kotlin, trait-method + UFCS resolution for Rust)
RAM-first pipeline: LZ4 compression, in-memory SQLite, single dump at end. Memory released after.
Distribution & operation
Native runtime set, zero infrastructure services: SQLite-backed, persists to ~/.cache/codebase-memory-mcp/
Auto-sync: Background watcher detects file changes in projects registered by active MCP sessions, including successful explicit index_repository targets, and re-indexes them automatically.
Route nodes: REST endpoints are first-class graph entities
Available on: npm, PyPI, Homebrew, Scoop, AUR, go install
Indexing modes
Mode
Use it for
full
Full indexing with semantic analysis. This is the default.
moderate
A filtered index with semantic analysis.
fast
A filtered index without semantic analysis.
cross-repo-intelligence
Link services across indexed projects, optionally limited by target_projects.
Team-Shared Graph Artifact
Share a compressed graph snapshot so teammates can reuse an indexed baseline.
.codebase-memory/graph.db.zst is a zstd-compressed snapshot of the knowledge graph that lives next to your source. The artifact is opt-in: index_repository writes it only with persistence: true (the default is false), and once it exists later indexes refresh it. Local indexes persist regardless. When a teammate indexes a clone with an existing artifact, CBM imports that baseline and applies their local changes incrementally.
Format: SQLite database, indexes stripped, VACUUM INTO compacted, then zstd 1.5.7 compressed (8 to 13:1 ratio typical)
Two tiers:
Best (zstd -9 + index strip + VACUUM INTO); written by index_repository with persistence: true
Fast (zstd -3); refreshes an existing artifact on every other index, including the watcher's low-latency incremental updates
Bootstrap: when no local DB exists but the artifact is present, index_repository imports the artifact first, then runs incremental indexing, reusing the shared baseline where possible
Merge policy: the first export creates .codebase-memory/.gitattributes with merge=ours. Check your repository's Git merge-driver setup before relying on it for concurrent artifact updates.
Commit it deliberately: the artifact is rewritten on every index, including the watcher's Fast tier, and git stores each rewrite as a full new blob. Committing every refresh is what turns a 20 MB file into gigabytes of history; one team reached ~6 GB across ~350 commits of this single path. Pick a cadence (a release, a milestone, a nightly job) rather than committing every save.
Git LFS, if it must move on every commit: track it from the repo-root.gitattributes and leave the auto-created .codebase-memory/.gitattributes in place; the nearer file goes on supplying merge=ours, and only filter comes from the root:
Track only the .zst; artifact.json is small and carries the schema version. The attribute applies to future commits only. Removing existing blobs requires a separate history migration. Two costs to weigh before adopting it: GitHub meters LFS storage and bandwidth, and its objects cannot be pruned without contacting support; and every teammate needs git lfs install; without it their checkout leaves a pointer file where the artifact should be, the integrity-checked import refuses it, and they fall back to a full reindex.
Optional: shared export is opt-in, and Git commits remain your choice. Add .codebase-memory/ to .gitignore if you prefer everyone to reindex from scratch.
The artifact provides a shared baseline while each developer keeps a local query database.
How It Works
codebase-memory-mcp is a structural analysis backend: it builds and queries the knowledge graph. The backend does not require a generative model. It relies on your MCP client (Claude Code, or any MCP-compatible agent) to be the intelligence layer.
code
You: "what calls ProcessOrder?"
Agent calls: trace_path(function_name="ProcessOrder", direction="inbound")
codebase-memory-mcp: executes graph query, returns structured results
Agent: presents the call chain in plain English
With MCP, your coding agent translates a question into tool calls. CBM Lens can also use an optional browser model for explanations; this does not change how the backend indexes or queries code.
Performance
Published backend benchmarks on Apple M3 Pro. These are separate from the CBM Lens checks:
Operation
Time
Notes
Linux kernel full index
3 min
28M LOC, 75K files → 4.81M nodes, 7.72M edges
Linux kernel fast index
1m 12s
1.88M nodes
Django full index
~6s
49K nodes, 196K edges
Cypher query
<1ms
Relationship traversal
Name search (regex)
<10ms
SQL LIKE pre-filtering
Dead code detection
~150ms
Full graph scan with degree filtering
Trace call path (depth=5)
<10ms
BFS traversal
Memory-aware pipeline: Indexing keeps hot data in memory and can spill extraction results to disk when needed. Memory is released back to the OS after indexing completes.
More focused context, fewer tokens: The project's 31-repository preprint reports 10× lower token use and 2.1× fewer tool calls compared with file-by-file exploration. These March 2026 backend measurements cover the study's workloads and agent setup; the paper documents the full evaluation. An earlier five-query comparison reported ~3,400 versus ~412,000 tokens for structural queries. The study's system-performance measurements include the Linux indexing and Cypher timings above.
Explore the measurement guide to evaluate quality, latency and token use on your own workload.
Troubleshooting & Diagnostics
codebase-memory-mcp indexes and queries locally and collects no telemetry. When you hit something we can't reproduce on our side (a slow memory climb over hours, a performance regression, a leak that only appears after days of real use), we have no diagnostic data unless you choose to send it. Here is how to capture it yourself.
Capture a diagnostics log
Set CBM_DIAGNOSTICS=1 before the first daemon-backed MCP session starts, then reproduce the problem (let it run as long as it takes; a slow leak needs time to show in the trend). The shared daemon captures this setting from the session that starts it. If it is already running, close all daemon-backed sessions so it exits before changing the setting. The daemon creates a fresh owner-private cbm-diagnostics-<pid>-<random> directory below the system temp directory ($TMPDIR or /tmp on macOS/Linux, %TEMP% on Windows). The exact paths are recorded by the diagnostics.start event in ${CBM_CACHE_DIR}/logs/cbm-daemon.log:
File
What it is
trajectory.ndjson
The memory trajectory: one JSON line every 5 s with rss, committed (Windows commit charge), peak_*, page_faults, fd, and queries. This is the file we need for memory/leak reports: the trend over time is what pinpoints a leak. It is kept on disk after the server exits (so you can grab it post-mortem) and rotates to trajectory.ndjson.1 past ~8 MB.
snapshot.json
The latest snapshot only; handy for a quick live check. Removed on clean exit.
The private randomized directory prevents another local account from pre-placing a link or special file at a predictable diagnostics path. Its <pid> component is the shared daemon's process ID, also recorded by the daemon.start event. Set the variable consistently in the env block of each agent's MCP server config, or export it before launching the first session.
What to share
When you open a memory/performance issue, attach the .ndjson trajectory: it contains no source code or query text, only resource counters. If you'd rather not attach a file, paste it (or an agent's summary of it) into the issue: your assistant can read the NDJSON directly and report whether rss/committed grow monotonically, how fast, and relative to query count; which is exactly what we need to find the cause.
Installation
Pre-built Binaries
Platform
Archive
macOS (Apple Silicon)
codebase-memory-mcp-darwin-arm64.tar.gz
macOS (Intel)
codebase-memory-mcp-darwin-amd64.tar.gz
Linux (x86_64)
codebase-memory-mcp-linux-amd64.tar.gz
Linux (ARM64)
codebase-memory-mcp-linux-arm64.tar.gz
Windows (x86_64)
codebase-memory-mcp-windows-amd64.zip
Every release includes checksums.txt with SHA-256 hashes. The executable is self-contained; no adjacent data file is required. Linux -portable archives contain the fully static builds; ordinary platform archives use their native system ABI.
Windows note: SmartScreen may show a warning for unsigned software. Click "More info" → "Run anyway". Verify integrity with checksums.txt.
The flake exposes two server packages plus the standalone frontend:
Package
Contents
default (codebase-memory-mcp)
Standard server, no UI
codebase-memory-mcp-ui
Server with CBM Lens embedded (--ui=true works)
graph-ui
Just the built CBM Lens assets (dist/)
Run directly without installing:
bash
# Standard server
nix run github:DeusData/codebase-memory-mcp
# Server with CBM Lens embedded
nix run github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui -- --ui=true --port=9749
# then open http://127.0.0.1:9749
Or build a binary into ./result/bin/codebase-memory-mcp:
bash
nix build github:DeusData/codebase-memory-mcp # standard
nix build github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui # with UI
Working in a clone? Use . in place of the flake URL, e.g. nix run .#codebase-memory-mcp-ui -- --ui=true, or drop into a shell that puts the binary on PATH with nix shell .#codebase-memory-mcp-ui.
Note: launched by hand (not from an MCP client) the server exits as soon as stdin closes; that's normal MCP behaviour. Keep stdin open while testing the UI, e.g. sleep infinity | codebase-memory-mcp --ui=true --port=9749. The codebase-memory-mcp-ui package embeds the UI at build time; nix run'ing the standard default package with --ui=true will refuse to start the HTTP server.
Install via Claude Code
code
You: "Install this MCP server: https://github.com/DeusData/codebase-memory-mcp"
Install via Claude Code plugin
This repository is also a Claude Code plugin marketplace, so you can find and
add the server from inside Claude Code:
bash
claude plugin marketplace add DeusData/codebase-memory-mcp
claude plugin install codebase-memory@codebase-memory
The plugin registers the MCP server only. It starts the server with
npx -y codebase-memory-mcp@<version>, pinned to the current release; the first
start downloads that release's binary through the npm package, which verifies
it against the release checksums. The skill, the three graph agents and the
context hooks are not part of the plugin: they come from
codebase-memory-mcp install.
Use one or the other. install registers the same server for Claude Code, so
with both you run it twice, under two tool namespaces
(mcp__codebase-memory-mcp__* and
mcp__plugin_codebase-memory_codebase-memory-mcp__*). Choose install for the
full Claude Code setup; choose the plugin to add just the server without
running an installer.
Build from Source
Prerequisites: C/C++ compiler, zlib, and Node.js for CBM Lens
Contributor checks need a few extra tools.scripts/test.sh requires zip, python3, and curl. scripts/lint.sh requires cppcheck, clang-format, and clang-tidy; scripts/lint.sh --ci, which is what CI runs, skips clang-tidy.
bash
git clone https://github.com/DeusData/codebase-memory-mcp.git
cd codebase-memory-mcp
scripts/build.sh --with-ui # the shipped composition (CBM Lens embedded, needs Node.js)
scripts/build.sh # without the UI (development only)# Binary at: build/c/codebase-memory-mcp (codebase-memory-mcp.exe on Windows)
The --with-ui build embeds CBM Lens and the agent integration templates into the native executable. You do not need Node.js to run that executable. Optional browser model weights are downloaded separately after opt-in.
Tests
Counts on main at 4d9592b5, where CBM Lens arrived. Native and Go counts use the macOS configuration; platform-specific registrations differ.
Suite
Count
Checked for this README
Native C
9,388 tests across 189 suites
9,376 passed, 11 skipped, 1 failed on macOS (regex_compile_refuses_overlong_pattern)
UI unit tests (Vitest)
3,970 tests in 310 files
All passed
Go package wrapper
30 top-level tests
Counted in the source
npm package wrapper
24 tests
Counted in the source
Python package wrapper
43 tests
Counted in the source
Run the native test suite:
bash
scripts/test.sh # full: clean sanitizer build + all suites + guards
scripts/test.sh --suites <name> # one suite, incremental, seconds
build/c/test-runner --list-suites # what is available
Frontend development and validation:
bash
cd graph-ui
npm ci
npm run test:unit
npm run build
npm run dev
The dev server listens on port 5173 and proxies /api and /rpc to port 9749. Start the matching CBM backend first.
Package wrappers have their own test commands: go test ./... in pkg/go, npm test in pkg/npm, and python3 -m unittest discover -s tests -p 'test_*.py' in pkg/pypi. The Go wrapper requires Go 1.26.1 or newer.
scripts/test.sh is the C test entry point used by CI. The canonical local artifact-flow check builds both stripped/unstripped candidates, defaults to the stripped candidate for this explicitly unscanned local run, packages those exact bytes, extracts the archive, and smokes it:
scripts/package-release.sh is intentionally a lower-level immutable boundary:
it accepts only an already-final --selected-binary plus its
--expected-sha256; it never builds, strips, signs, or relinks the executable.
Manual MCP Configuration
If you prefer not to use the install command
Add to ~/.claude.json (user scope) or project .mcp.json:
Restart your agent. Verify with /mcp; you should see codebase-memory-mcp with 17 tools.
Multi-Agent Support
install configures 46 supported automatic/conditional client surfaces: 40 detected automatically and 6
conditional or explicit. “Conditional” means the installer writes only when the
documented platform or an explicit, already-existing config path proves the
target is active. It never flips experimental feature flags, enables plugins,
YOLO modes, global permission bypasses, or third-party instruction trust.
Where a client has a documented custom-agent format, the installer creates three
exact-owned definitions from one canonical contract:
Scout (Tier 1): about 3 to 4 narrow calls for fast positive, provisional discovery; no absence, exhaustive-impact, or dead-code claims.
Verify (Tier 2, default): task-directed graph evidence, exact source checks, path coverage for every cited file, and scope coverage before negative claims.
Auditor (Tier 3): bounded scope, current index generation, complete relevant pagination, broader relationship checks, and explicit unresolved limitations.
Every direct tier batches check_index_coverage for its evidence paths and reads
flagged ranges or skipped/excluded files directly. A clean coverage result means
only “no recorded gap,” never proof of completeness. Clients without safe child
MCP access receive the same three tiers as parent-handoff agents; the parent must
supply project, generation, pagination state, graph evidence, and coverage
results. Updates migrate only byte-identical prior Verify definitions and never
overwrite user-modified agents.
Agent
Activation
MCP config
Durable context / augmentation
Claude Code
Detected
~/.claude.json
Skill + three exact-tool graph agents; SessionStart, SubagentStart, non-blocking PreToolUse for Grep/Glob/Bash, post-Read coverage, and a drift note after repeated searches without a graph call
Codex CLI
Detected
$CODEX_HOME/config.toml
Managed AGENTS.md activation pointer, ~/.agents/skills/codebase-memory/SKILL.md, three read-only agents; SessionStart + SubagentStart, PreToolUse graph lookup for Bash searches
Gemini CLI
Detected
.gemini/settings.json
GEMINI.md, three explicit read/graph-tool subagents; BeforeTool, AfterToolread_file coverage, and SessionStart
Zed
Detected
platform settings.json (JSONC)
AGENTS.md + shared skill
OpenCode
Detected
$OPENCODE_CONFIG or resolved global config
AGENTS.md, skill, three deny-by-default read-only agents; plugin adds grep/glob graph lookup, post-read coverage, first-tool-result session context, and post-compaction reinjection on OpenCode 1.18.x
Antigravity
Detected
.gemini/config/mcp_config.json
.gemini/GEMINI.md
Aider
Detected
none
CONVENTIONS.md via .aider.conf.yml
KiloCode
Detected
.config/kilo/kilo.jsonc
Rule + three graph-tool subagents with deny-by-default permissions
Steering, skill, three JSON agents with isolated Scout/Analysis-profile MCP and explicit graph-tool selectors (includeMcpJson: false)
Junie
Detected
.junie/mcp/mcp.json
Skill + three graph subagents for EAP-capable builds; Scout and Analysis server aliases hard-limit the tier tool surfaces; no ineffective EAP SessionStart hook
AGENTS.md, skill, three droids with exact per-tier graph-tool lists (without additive whole-server exposure); SessionStart + post-Read coverage on macOS/Linux, withheld on Windows
Crush
Detected
.config/crush/crush.json
Managed context path with explicit parent-to-child handoff
Goose
Detected
.config/goose/config.yaml
.goosehints
Mistral Vibe
Detected
$VIBE_HOME/config.toml
AGENTS.md, skill, and three matched agent/prompt pairs with explicit read-only graph-tool allowlists
Grok Build
Detected
$GROK_HOME/config.toml
Owned rules/codebase-memory.md, skill, three graph agents with named-server mcpInheritance and exact server__tool dispatcher ids; context hooks withheld because its passive hook events discard stdout
Qoder CLI
Detected
~/.qoder/settings.json
Skill, three directly MCP-attached agents with named-server scoping and exact per-tier graph-tool lists; SessionStart, SubagentStart, and post-Read coverage, including documented PowerShell execution on Windows
Kimi Code CLI
Detected
$KIMI_CODE_HOME/mcp.json (default ~/.kimi-code)
Same-root AGENTS.md + skill; fail-open UserPromptSubmit hook in config.toml
GitLab Duo CLI
Detected
$GLAB_CONFIG_DIR/duo/mcp.json or platform fallback
Fail-open user SessionStart on macOS/Linux; hook withheld on Windows; no experimental global skill enablement
Rovo Dev CLI
Detected
configured override or ~/.rovodev/mcp.json
Global AGENTS.md, skill + three read-only handoff subagents; no undocumented hook
Amp
Detected
~/.config/agents/skills/codebase-memory/mcp.json
Colocated skill + ~/.config/amp/AGENTS.md; no plugin
Devin CLI / Local
Detected
~/.config/devin/config.json (platform app-data path on Windows)
Same-root AGENTS.md + skill; macOS/Linux UserPromptSubmit + PostCompaction, and SessionStart only when Claude does not already provide it; hooks withheld on Windows
Tabnine
Detected
~/.tabnine/mcp_servers.json
MCP only; no experimental/YOLO setting
Continue / cn
Conditional
Existing ~/.continue/config.yaml or $CBM_CONTINUE_CONFIG_PATH
MCP only
Visual Studio
Conditional, Windows
~/.mcp.json
MCP only
TRAE
Conditional
Existing $CBM_TRAE_CONFIG_PATH
MCP only
Roo Code
Conditional
Existing $CBM_ROO_CONFIG_PATH
MCP only
Amazon Q Developer IDE
Detected
~/.aws/amazonq/default.json (preserves an existing agents/default.json or legacy mcp.json)
MCP only
CodeBuddy Code CLI
Detected
~/.codebuddy/.mcp.json (preserves an active deprecated/legacy file)
CODEBUDDY.md, skill, three read-only graph agents; beta hooks are not auto-installed
IBM Bob Shell
Detected by bob
~/.bob/mcp_settings.json
Shared rule; no invented hook or agent
Pochi
Detected
~/.pochi/config.jsonc (mcp)
README.pochi.md, skill, and three readFile-only parent-handoff agents
Pi
Detected
none
~/.pi/agent/AGENTS.md + skill; MCP/subagents require an explicit reviewed extension
IBM Bob IDE
Conditional
Existing ~/.bob/mcp.json
Shared rule + IDE skill; no invented hook or agent
Skill and three direct-MCP graph-tool subagents (Scout/Verify/Auditor); preserves user AGENTS.md
agentty
Detected
~/.agentty/mcp.json (overridden by $AGENTTY_MCP_CONFIG)
Skill in ~/.agentty/skills (Claude-compatible SKILL.md)
Sourcegraph Cody
Explicit opt-in
Existing $CBM_CODY_CONFIG_PATH
MCP only
For Codex, install keeps only a tiny managed activation pointer in global
$CODEX_HOME/AGENTS.md; all detailed behavior lives in the installed codebase-memory skill.
Fresh installs create the pointer, upgrades replace the legacy full managed block while preserving
all user-owned bytes, and uninstall removes only the managed pointer.
Skill placement follows the cross-client user skill root and does not follow
CODEX_HOME. On upgrade, the installer migrates an installer-owned legacy
$CODEX_HOME/skills/codebase-memory/SKILL.md only after verifying the canonical
copy. Modified or foreign files are preserved; --force explicitly authorizes
replacement before migration.
Sessions, compaction, and subagents
Hooks installed by this project are fail-open and context-only. Claude Code's
PreToolUse observes Grep/Glob/Bash and injects matching graph symbols as
additionalContext; PostToolUse on Read adds targeted coverage context when
the graph could not fully parse or index that file. It never denies or replaces
the requested tool call. Codex CLI's PreToolUse on Bash gives its shell
searches the same context. GitHub Copilot CLI and VS Code's Copilot runtime get
it after the search: their preToolUse can only decide permissions, so a
PostToolUse entry for Bash, Grep and Glob appends the graph symbols to
the tool result. In Claude Code, once a session has run three searches with no
codebase-memory-mcp tool call, the next augmentation adds one line saying so;
the next note comes after six searches, then twelve. Claude's PostToolUse
entry also matches the graph tools, and a graph tool call resets the count.
Claude Code, Codex CLI, Qwen Code, GitHub Copilot CLI, and VS Code's Copilot
runtime receive paired session/subagent context where the vendor exposes a
documented context-output contract. Codex users must review and trust installed
hooks through /hooks; changing a hook definition changes its trust hash, so an
update can require re-trust. Qoder uses SessionStart, SubagentStart, and
post-Read coverage, including its documented PowerShell executor on Windows.
Kimi uses UserPromptSubmit, while Hermes uses pre_llm_call; both retain their
documented Windows execution paths. Devin installs
UserPromptSubmit and PostCompaction on macOS/Linux and adds SessionStart
only when Claude's equivalent managed hook is not present. GitLab Duo gets a
narrowly scoped macOS/Linux user SessionStart entry on its experimental hook
surface. GitLab Duo, Devin, and Factory hooks are withheld on Windows
because those vendors do not document a deterministic shell/executor contract
there. Gemini CLI, Factory Droid, and Augment also add documented post-read/view
coverage context but expose no equivalent documented child-start context.
For runtimes without a stable context-producing lifecycle event, durable files
carry the contract across fresh sessions and compaction: verify the graph project
and index freshness, query structural facts in the parent, then pass the project,
qualified symbols, paths, and call-chain evidence in every delegated task.
Claude, Codex, Gemini, Kiro, Qwen, Copilot, CodeBuddy, OpenCode, Kilo, Vibe,
Qoder, Junie, Factory, and Grok Build receive Scout, Verify, and Auditor graph profiles.
Kiro embeds this MCP server with --tool-profile scout for Scout and
--tool-profile analysis for Verify/Auditor. Junie registers equivalent named
server aliases because its subagent schema filters by server rather than by
individual tool. Both process profiles use positive allowlists: Scout exposes
seven fast inspection tools, Analysis exposes eleven, and future or mutating
tools remain unavailable until explicitly reviewed. If either Junie alias
collides with user configuration, the installer preserves it and installs
parent-handoff profiles instead. Qoder combines its documented named-server
selection with exact tier-specific MCP tool IDs. Factory uses exact registered
MCP tool IDs without its additive mcpServers field, which would expose the
whole server. Codex, Kilo, Vibe, and other capable formats likewise enumerate
the narrowest supported tool set. Rovo, Cursor, Augment, Pochi, and Cline use parent handoff where direct
child MCP is unavailable or unsafe; Pochi is limited to readFile, and Cline
child agents cannot use MCP.
Cline's file hooks auto-activate when present, and current Cline does not
reliably consume their context output, so automatic adapters are withheld and
older owned adapters are cleaned up. CodeBuddy's beta, version-gated hooks are
not auto-installed. Junie's EAP
SessionStart output is documented as ignored, so no context hook is installed.
Junie custom agents remain EAP-dependent. Qoder can resolve higher-priority
project or plugin agents before user agents with the same name; reload the
client after installation or profile changes.
Cursor context
hooks are withheld: session context injection has a known race, subagentStart
is control-only, and read-only subagents cannot safely receive MCP access. Grok Build's
passive hook events (SessionStart, SubagentStart, PostToolUse) discard
stdout and PreToolUse honors only deny/rewrite decisions, so its context hooks
are withheld; Grok also reads Claude and Cursor MCP, skill, and hook files
through its compat layer, and the native config.toml entry shadows that copy
by name. Rovo
has no documented session context-output hook, and Bob
documents neither a suitable hook nor a custom-agent surface. Those surfaces are
not approximated with invented augmentation. Kimi plugins, Amp plugins, and
GitLab experimental global skills remain opt-in.
OpenClaw reinjects the Codebase Knowledge Graph (codebase-memory-mcp) AGENTS
section after compaction and places the same guidance in TOOLS.md, the bootstrap
files inherited by its subagents. Automatic augmentation covers the active/default
workspace. Separate agents.list[].workspace directories require making that
workspace active for installation or copying the managed block there.
The installed Claude shim is named cbm-code-discovery-gate for backward
compatibility; despite the legacy name, it never gates or blocks.
Manual or UI-managed integrations
These are intentionally not counted as automatic installs: Qodo MCP is added
through its UI and may be governed by enterprise allowlists; Warp MCP is managed
through Warp Drive/UI or per invocation (only the shared skill is automatic);
JetBrains AI Assistant / ACP is IDE-managed; GitHub Copilot coding agent, Jules,
and CodeRabbit are cloud/repository-managed; Replit exposes a remote/service
integration rather than a stable local user-global client; BLACKBOX AI does not
document a stable arbitrary user-global MCP/instruction/agent schema; Plandex has
no stable global registry safe to mutate; and SWE-agent uses explicit YAML and is
no longer a suitable automatic global target.
CLI Mode
Every MCP tool can be invoked as a local, one-shot command. CLI tools connect to the shared coordination daemon (starting one if none is already running) to get a crash-safe exact-build admission lease for the command's lifetime; they never register with the background watcher, and a daemon a command had to start itself exits again once the command's connection closes and no other session is using it. index_repository is the only exception internally: it starts a temporary, exact-build supervised worker for the index, then stops that worker before the CLI command exits; the worker holds its own lease until exit.
Commands that mutate graph data use shared OS-backed, per-project locks. This serializes conflicting work from CLI and MCP sessions on the same project while allowing unrelated projects to proceed independently.
When stderr is an interactive terminal, the CLI automatically shows lifecycle and indexing progress. Pass --progress to force the same feedback when stderr is redirected or the command is run non-interactively. Pass --quiet to disable automatic terminal progress and ordinary diagnostics while retaining errors; it cannot be combined with --progress or outer cli --verbose. Routine informational logs are quiet by default; pass outer cli --verbose to include them. Progress and logs use stderr while stdout remains reserved for the command result. Read tools return a compact tree by default; pass a tool's --format json for machine-readable payload JSON, or outer --json for the full MCP envelope.
Large compact-tree tables may start with a response-local <section>_refs directory and an explicit <section>_ref_rule. A cell such as @0+handler.go reconstructs to ref 0's prefix plus handler.go. References are local to that sibling <section> table and expansion is non-recursive: entries inside <section>_refs are always literal prefixes. This is limited to declared path and qualified-name columns and activates only when the exact rendered table is at least 15% and 64 bytes smaller and a conservative model-neutral token-shape proxy also improves by at least 1%. Search and trace likewise render direct and prefix-grouped tree shapes and keep the smaller complete representation, so singleton or scattered answers do not pay directory overhead. Keys are declared once per table but never cryptically abbreviated, and --format json keeps stable literal strings for machine consumers. Both gates are deterministic; exact token counts still depend on the caller's tokenizer.
Lean responses truncate semantically, never by cutting arbitrary bytes from code or identifiers. Ranked graph rows are retained ahead of raw grep rows and diagnostic summaries; omitted rows/sections report totals, has_more, and a strictly advancing continuation offset or cursor. If even the first whole row cannot fit, CBM asks for a higher budget and emits no self-looping cursor. max_output_tokens is model-neutral sizing guidance: CBM enforces a deterministic ceiling of four UTF-8 bytes per requested token, so it is not a tokenizer-exact count. Detail flags such as diagnostics, source_mode, and detail opt into heavier fields. search_code pages ranked rows with result_limit/result_offset (limit remains a compatibility alias), raw rows with raw_limit/raw_offset, and directory summaries with directory_limit/directory_offset. Raw lines default to a UTF-8-safe match-centered preview; each row reports content_start_byte, returned/total byte counts, match byte bounds when known, and a content continuation offset. Pass raw_content_offset to page the original line without moving the raw-row cursor. match_limit and source_max_lines bound per-result details, with exact omission metadata. detect_changes pages changed files, impacted symbols, and module summaries independently; prefer its snapshot-bound *_cursor continuations, which reject changed commits, worktree bytes, graph generation, or semantic arguments instead of silently skipping or duplicating rows.
Every response is standard UTF-8. Identifiers, paths, and raw search previews preserve POSIX byte-string identities: a preserved value containing malformed UTF-8 is emitted reversibly as @bytes:<lowercase hex of every original byte>. A valid preserved value that literally begins with the reserved @bytes: or @utf8: prefix is emitted as @utf8:<original value>, so decoding is unambiguous: strip one @utf8: prefix for literal UTF-8, or hex-decode one @bytes: prefix for original bytes. Ordinary valid UTF-8 is unchanged and pays no output-token overhead. To keep code readable, source bodies replace malformed UTF-8 with U+FFFD; use the pageable raw search preview when byte-exact source inspection is required.
Use cli <tool> --help to see the flags generated from that tool's input schema:
bash
codebase-memory-mcp cli index_repository --repo-path /path/to/repo
codebase-memory-mcp cli list_projects
# Use the "name" returned by list_projects as the project value.
codebase-memory-mcp cli search_graph --project my-project --name-pattern '.*Handler.*' --label Function
codebase-memory-mcp cli trace_path --project my-project --function-name Search --direction both
codebase-memory-mcp cli query_graph --project my-project --query 'MATCH (f:Function) RETURN f.name LIMIT 5'# Force human-readable progress without contaminating stdout.
codebase-memory-mcp cli --progress index_repository --repo-path /path/to/repo
# Suppress automatic terminal progress and non-error diagnostics.
codebase-memory-mcp cli --quiet list_projects --format json
codebase-memory-mcp cli search_graph --project my-project --label Function --format json
codebase-memory-mcp cli list_projects --format json --detail stats | jq '.projects[].name'
Tools requiring arguments accept JSON piped on stdin. Tools with only optional arguments, such as list_projects, accept flags, --args-file, or redirected JSON files without waiting on an inherited pipe. Inline JSON remains accepted for backward compatibility but is deprecated in favor of flags, --args-file, or stdin.
MCP Tools
Indexing
Tool
Description
index_repository
Index a repository into the graph. A successful call from an active MCP session registers the target for auto-sync while that session remains connected (under daemon start, also after it disconnects), subject to the global auto_watch setting; CLI calls do not register watches. Waits for the whole index by default; pass async: true to start it in the daemon and return at once, then poll with status: true (see below).
list_projects
List indexed projects with paging; use detail: "stats" for node/edge counts and database sizes.
delete_project
Remove a project and all its graph data, including its ADR sidecar (<db>.adr.db).
index_status
Check indexing status of a project.
check_index_coverage
Check whether exact paths or a scope are indexed and fresh. A clean result means no recorded gap, not proof of completeness.
Long indexes and client call deadlines. A synchronous index_repository on a large
repository can take longer than an MCP client allows one tool call (some IDE clients give up
after a fixed deadline). When a client cancels or disconnects, the daemon cancels an index that
nobody else is waiting for, so retrying the same blocking call never finishes. Use the async
mode instead:
index_repository(repo_path="/abs/path", async: true) starts the index in the daemon (or
joins the one already running for that project) and returns immediately with
state (queued/running). The job keeps running even if the client cancels, times out or
disconnects; only stopping the daemon ends it.
index_repository(repo_path="/abs/path", status: true) reports state
(queued, running, cancelling, succeeded, failed, cancelled), started_at,
finished_at and an error summary. Poll it until the state is succeeded, failed or
cancelled. Pass the same repo_path (and name, if the index call used one).
async and status are exclusive; async does not apply to cross-repo-intelligence. Both
need the daemon-backed server (the default codebase-memory-mcp entry point). A temporary
daemon (started on demand rather than by codebase-memory-mcp daemon start) stops, and
cancels its jobs, when its last client disconnects. A connected MCP session keeps it alive, so
async from your editor works. A one-shot codebase-memory-mcp cli index_repository --async
that is the daemon's only client is refused with a clear error, because the job would die the
moment the command exits: run codebase-memory-mcp daemon start first, keep an MCP session
open, or call without --async. status works everywhere. When a synchronous call was cut
short, the next
index_repository or status call for that project carries a notice suggesting the async
mode. index_status keeps describing the published graph and its freshness.
Querying
Tool
Description
search_graph
Structural, BM25, and semantic search. Page structural rows with offset/limit and ranked semantic rows independently with semantic_offset/semantic_limit.
trace_path
Trace callers, callees, data flow, or cross-service paths (alias: trace_call_path). Depth 1 to 15; tests are excluded by default. mode: calls (default), data_flow, cross_service, topology (Kubernetes objects, routes, controllers and what they deploy; start at a Resource or Route), deploy (what a CI job, script or Make rule runs and deploys; outbound from the job, inbound from a chart), run (what a Service or workload runs: workload → Image → Dockerfile → the program its container starts, with the environment each container is given; outbound from a Service, workload or compose service).
detect_changes
Map git diff to affected symbols + blast radius with risk classification.
query_graph
Execute Cypher-like graph queries (read-only).
get_graph_schema
Node/edge counts, relationship patterns, property definitions per label. Run this first.
compare_graphs
Compare two indexed snapshots: node/edge additions and removals between a base and a target.
get_code_snippet
Read source code for a function by qualified name.
get_file_outline
Declaration outline of one repository-relative file in source order, with optional label filter and paging.
Grep-like text search within indexed project files.
manage_adr
CRUD for Architecture Decision Records (get reads, update replaces the whole document, set_sections rewrites only the named sections and leaves every other byte untouched, sections lists headings). Query modes do not wait behind a same-project reindex; writes remain serialized.
ingest_traces
Store observed call paths or caller/callee pairs under a run label. Both endpoints must resolve to indexed symbols; unresolved names are reported.
manage_adr(mode='set_sections') writes one or more sections by name and splices them into the stored document, so text outside the named sections (including a preamble, code fences and section ordering) is preserved byte-for-byte. Any ## Heading works, not just the conventional PURPOSE / STACK / ARCHITECTURE / PATTERNS / TRADEOFFS / PHILOSOPHY set; names match exactly, including case. Writing the same section twice is a no-op, so a retry after a lost response cannot duplicate content.
manage_adr query modes (get and sections) use the server's cached query store so they can proceed while a same-project reindex is running. If another process publishes a replacement store during reindexing, they can return the pre-publication ADR until idle eviction refreshes that cache. Updates remain serialized through the project mutation guard.
trace_path exposes the start nodes' stored self_recursive (direct recursion)
and recursive (direct recursion or a call-graph cycle) signals in both tree and
JSON output. Across multiple start nodes, a flag is true if any records true,
false if all record false, and null otherwise (including missing legacy
metadata). These are static-analysis signals, not counts of resolved self-calls.
Static CALLS indexing omits direct self-call edges; traces that include CALLS state
self_call_policy: omitted. Caller/callee totals still count graph-reachable
nodes, so a recursive function can have zero callees. Recursion metadata remains
visible on continuation pages and when an output budget prevents rows from fitting.
MENTIONS links a documented definition to the code its doc comment references (C# <see cref>, <seealso cref>, <exception cref>, <inheritdoc cref>), one edge per pair with via, syntax, tier (exact or unique), first line and count. A reference is bound only when it names exactly one definition under the language's own lookup rules; the rest are counted by reason in index_status under doc_links (samples with diagnostics='full').
Qualified Names
get_code_snippet uses qualified names: <project>.<path_parts>.<name>. Use search_graph to discover them first.
Supported Cypher (openCypher read subset)
query_graph is a read-only openCypher subset:
Clauses: MATCH, OPTIONAL MATCH, multiple MATCH, WHERE, WITH (+ WITH … WHERE), RETURN, ORDER BY, SKIP, LIMIT, DISTINCT, UNWIND, UNION / UNION ALL, CASE.
WHERE: = <> < <= > >=, AND/OR/XOR/NOT, IN, CONTAINS, STARTS WITH, ENDS WITH, IS [NOT] NULL, regex =~, label test n:Label, and EXISTS { (n)-[:TYPE]->() } (single-hop existence; great for dead-code, e.g. WHERE NOT EXISTS { (f)<-[:CALLS]-() }).
See docs/cbmignore.md for the full .cbmignore how-to: syntax, precedence across the ignore layers, and negation semantics.
Configuration
bash
codebase-memory-mcp config list # show all settings
codebase-memory-mcp config set auto_index true# auto-index on session start
codebase-memory-mcp config set auto_index_limit 50000 # max files for auto-index
codebase-memory-mcp config set auto_watch false# remove/disable all background watches (default: true)
codebase-memory-mcp config set ignore_worktrees true# skip linked git worktrees when auto-indexing (default: false)
codebase-memory-mcp config set doc_candidates false# skip doc<->code candidate scoring on huge repos (default: auto = on for full, off for moderate/fast)
codebase-memory-mcp config set watcher_enabled false# stop the watcher thread entirely (default: true)
codebase-memory-mcp config set watch_non_git true# also poll non-git project roots (default: false)
codebase-memory-mcp config set index_max_files 250000 # optional per-index source-file limit
codebase-memory-mcp config set index_max_source_mb 16384 # optional per-ind