SocratiCode

"There is only one good, knowledge, and one evil, ignorance." — Socrates

Your AI reads code. SocratiCode understands it.

The open-source codebase context engine: give any AI instant automated knowledge of your entire codebase (and infrastructure) — at scale, zero configuration, fully private, completely free.

Kindly sponsored by Altaire Limited

🛡️ Need MCP governance together with codebase context? See our sibling project JanuScope — the local-first MCP policy proxy: tool blocking, SQL-mutation gate, PII redaction, audit, rate-limit.

If SocratiCode has been useful to you, please ⭐ star this repo — it helps others discover it — and share it with your dev team and fellow developers!

💬 Questions or just want to chat? Join us on Discord.

☁️ SocratiCode Cloud (private beta) — Hosted, shared team index built on the same engine as the open-source version, plus SSO, audit logs, branch-aware indexing, and VPC / air-gapped deployment options. The open-source core remains free forever. Request early access →

One thing, done well: deep codebase intelligence with zero setup, no bloat, and full automation. SocratiCode gives AI assistants deep semantic understanding of your codebase: hybrid search, cross-project search, polyglot code dependency graphs, symbol-level impact analysis and flow, interactive HTML graph explorer for visual navigation, and searchable context artifacts (database schemas, API specs, infra configs, architecture docs). Zero configuration: add it to an MCP host that supports local stdio servers, or use a supported plugin or extension. It manages everything automatically.

Production-ready, battle-tested on enterprise-level large repositories (up to and over ~40 million lines of code). Batched, automatic resumable indexing checkpoints progress — pauses, crashes, restarts, and interruptions don't lose work. The file watcher keeps the index automatically updated at every file change and across sessions. Multi-branch, multi-repo and multi-agent ready — multiple AI agents can work on the same codebase simultaneously, sharing a single index with automatic coordination and zero configuration.

Private and local by default — Docker handles everything, no API keys required, no data leaves your machine. Cloud ready for embeddings (OpenAI, Google Gemini) and Qdrant, and a full suite of configuration options are all available when you need them.

Code intelligence that belongs to you, AI and host agnostic — your codebase's understanding lives with the code, not locked to any one assistant, IDE or model. And because SocratiCode pre-computes the hard parts (blast radius, call-flow, dependency traversal), smaller models can handle architectural complex tasks that would otherwise need top-tier reasoning, saving even more on token cost.

The first Qdrant‑based MCP/Claude Plugin/Skill that pairs auto‑managed, zero‑config local Docker deployment with AST‑aware code chunking, hybrid semantic + BM25 (RRF‑fused) code search, polyglot dependency graphs with circular‑dependency visualisation, symbol‑level Impact Analysis (blast‑radius & call‑flow tracing across 18 languages), and searchable infra/API/database artifacts in a single focused, zero-config and easy to use code intelligence engine.

Benchmarked on VS Code (2.45M lines): SocratiCode uses 61% less context, 84% fewer tool calls, and is 37x faster than grep‑based exploration — tested live with Claude Opus 4.6. See the full benchmark →

Contents

• Quick Start
• Quick guides
• Plugins and host integrations
• Why SocratiCode
• Features
• Prerequisites
• Example Workflow
• Agent Instructions
• Configuration
• Language Support
• Ignore Rules
• Context Artifacts
• Environment Variables
• Docker Resources
• Testing
• Why Not Just Grep?
• FAQ
• Community
• SocratiCode Cloud
• License

Quick Start

Requirements: Node.js 18.17 or newer with npx on PATH, plus Docker running for the default local Qdrant and Ollama stack.

For one recommended setup path per host and short practical scenarios, use the quick guides. This README has the other installation paths and full configuration.

Quick install guidance for Claude Code, VS Code, and Cursor:

MCP hosts with a JSON mcpServers object can use this complete configuration:

{
  "mcpServers": {
    "socraticode": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"]
    }
  }
}

Configuration schemas are host-specific. Continue, VS Code, Zed, OpenCode, Gemini CLI, and Cline have dedicated examples in Plugins and host integrations.

Keeping SocratiCode up to date

SocratiCode has two independent update paths. The MCP engine is the socraticode package published to npm. Every npm-backed configuration below uses npx -y --prefer-online socraticode@latest, which checks npm for the current latest release whenever the MCP server starts. A running server cannot replace itself, and a newly published version can only be downloaded while the npm registry is reachable, so restart or reconnect the server after a release.

Native plugins and extensions also contain skills, instructions, manifests, or UI files. Update those through the host as shown below, then start a new session so the new plugin files load. Direct MCP installations contain only the engine and do not install SocratiCode's plugin skills.

@latest refers to npm's published latest distribution tag; it does not refer to a Git branch. --prefer-online forces npm to check for updated package metadata even when its cache is still fresh. If the same registry is temporarily unavailable, npm can still use an already populated cache; a first installation still requires registry access. See the npm exec cache documentation and npm distribution-tag documentation.

Restart your host. With the default local configuration, first use pulls the required Docker images and starts managed Qdrant. OLLAMA_MODE=auto reuses a detected native Ollama instance or starts managed Ollama, then downloads the local embedding model if it is not already available. Cloud and external embedding providers do not download a local model. Initial setup usually takes about five minutes, depending on the connection; later starts take seconds.

First time on a project: ask your AI: "Index this codebase". Indexing runs in the background; ask "What is the codebase index status?" to monitor progress. Depending on codebase size and whether you're using GPU-accelerated Ollama or cloud embeddings, first-time indexing can take anywhere from a few seconds to a few minutes (it takes under 10 minutes to first-index +3 million lines of code on a Macbook Pro M4). Once complete it doesn't need to be run again, you can search, explore the dependency graph, and query context artifacts.

Every time after that: just use the tools (search, graph, etc.). By default, server startup resumes the indexed project represented by the MCP process's working directory: a complete index gets its watcher and an incremental catch-up update, while interrupted indexing resumes from the last checkpoint. SOCRATICODE_AUTO_RESUME_PROJECTS and SOCRATICODE_AUTO_RESUME=all can select additional projects. A completed indexed project not handled at startup gets a fallback watcher start on its first search, status, or graph interaction. You can also explicitly start or restart the watcher with codebase_watch { action: "start" }.

Prefer a deliberate index snapshot? Set SOCRATICODE_WATCHER=off and SOCRATICODE_AUTO_RESUME=off for every MCP process that uses the checkout, then run codebase_update only when you want to refresh it. Existing indexes remain usable without rebuilding. Use SOCRATICODE_WATCHER=manual instead if explicit codebase_watch { action: "start" } should remain available. See Indexing Behaviour and Passing env vars by host.

Prefer updates at Git transitions? Set SOCRATICODE_WATCHER=git. Active indexed checkouts refresh when the checked-out ref or HEAD changes, without a native file watcher. See Git-triggered refresh.

macOS / Windows on large codebases: Docker containers can't use the GPU. For medium-to-large repos, install native Ollama (auto-detected, no config change needed) for Metal/CUDA acceleration, or use OpenAI embeddings for speed without a local install. Full details.

Recommended: For best results, add the Agent Instructions to your AI assistant's system prompt or project instructions file (CLAUDE.md, AGENTS.md, etc.). The key principle, search before reading, helps your AI use SocratiCode's tools effectively and avoid unnecessary file reads.

Claude Code users: If you installed the SocratiCode plugin, the Agent Instructions are included automatically as skills, so there is no need to add them to your CLAUDE.md. The plugin also bundles the MCP server, so you don't need a separate claude mcp add.

Advanced: cloud embeddings (OpenAI / Google), external Qdrant, remote Ollama, native Ollama, and dozens of tuning options are all available. See Configuration below.

Plugins and host integrations

SocratiCode can be installed as a native agent plugin, a VS Code editor extension, or a directly configured local stdio MCP server. These are separate integration types and use different configuration and update paths.

The SocratiCode engine requires Node.js 18.17 or newer with npx on PATH; some hosts require a newer Node.js version. The default local stack also requires Docker to be running. Docker is optional when Qdrant is external and embeddings use either a detected native Ollama instance or a cloud or external provider.

Claude Code plugin (recommended for Claude Code users)

The native plugin bundles the MCP server, workflow skills, and agent instructions. Install it for the current user:

claude plugin marketplace add giancarloerra/socraticode
claude plugin install --scope user socraticode@socraticode
claude plugin list

Start a new Claude Code session after installation. Existing sessions do not load a newly installed plugin. To enable automatic updates, open /plugin, select Marketplaces, select socraticode, and enable auto-update. To update manually:

claude plugin marketplace update socraticode
claude plugin update --scope user socraticode@socraticode

Run /reload-plugins or start a new session after updating. If SocratiCode was previously added as a standalone MCP server, remove that duplicate with claude mcp remove socraticode; the plugin already provides the server.

For custom providers or external Qdrant, put inherited variables in Claude Code's user settings and restart the session:

{
  "env": {
    "EMBEDDING_PROVIDER": "openai",
    "OPENAI_API_KEY": "<your key>"
  }
}

Keep this user-scoped file private, or provide secrets through the process environment. Never commit secret values.

Pinning the engine version

By default the plugin launches npx -y --prefer-online socraticode@latest, which checks npm for the current release every time the server starts. Each start resolves independently, so two servers begun at different moments on one host can run different releases.

SOCRATICODE_SPEC replaces the package specification. Unset, it resolves to socraticode@latest and the launcher is unchanged, so an existing installation needs no reconfiguration. Set it where the plugin's server inherits it, in Claude Code's user settings:

{
  "env": {
    "SOCRATICODE_SPEC": "socraticode@1.14.0"
  }
}

Remove the entry to go back to the default; an empty string is a value, not an absence, and would be passed to npm as the specification.

Every server on that host then resolves the same build. npm is still contacted for metadata, so start-up is not free, but a pinned version is already in the npx cache after its first resolution and is not downloaded again on each new release. A pinned engine no longer updates on its own — raise the value deliberately to pick up a release.

This applies to the Claude Code plugin only. Codex and Cursor read the shared .mcp.json, whose launcher is unchanged; the VS Code extension exposes the equivalent socraticode.command and socraticode.args settings.

See the Claude Code plugin documentation.

Claude Code MCP-only installation

For a user-scoped installation without the bundled skills:

claude mcp add --scope user socraticode -- npx -y --prefer-online socraticode@latest
claude mcp list

Start a new session, or run /mcp and select Reconnect. The explicit @latest command checks for the latest published engine each time the server starts. claude mcp add defaults to project-local scope when --scope user is omitted. See the Claude Code MCP documentation.

OpenAI Codex plugin

The Codex plugin bundles SocratiCode's MCP server, skills, and instructions. Add the Git marketplace and install the plugin for the current user:

codex plugin marketplace add giancarloerra/socraticode --ref main
codex plugin add socraticode@socraticode
codex plugin list --available --json

Start a new Codex task or CLI session after installation. Reopening an existing task does not load newly installed skills or MCP tools. Update and verify with:

codex plugin marketplace upgrade socraticode
codex plugin add socraticode@socraticode
codex plugin list --available --json

If SocratiCode fails to start in Codex Desktop with No such file or directory, Codex may not be able to find npx. Run command -v npx. Some desktop launches use only /usr/bin:/bin:/usr/sbin:/sbin; if npx is installed elsewhere, add its directory to PATH in the custom Codex configuration below, then restart Codex.

Codex has a public plugin directory, but SocratiCode is not currently published there. Plugins are supported in the ChatGPT desktop Codex environment and Codex CLI. The Codex IDE extension supports shared MCP configuration, not plugin discovery. See the OpenAI plugin documentation.

Codex currently exposes enablement and tool policy for a plugin's bundled MCP server, but not a documented per-user environment override for that bundled definition. To keep the plugin skills while using custom SocratiCode variables, disable only the bundled server and add one top-level server in ~/.codex/config.toml:

[plugins."socraticode@socraticode".mcp_servers.socraticode]
enabled = false

[mcp_servers.socraticode]
command = "npx"
args = ["-y", "--prefer-online", "socraticode@latest"]

[mcp_servers.socraticode.env]
QDRANT_MODE = "external"
QDRANT_URL = "https://xyz.qdrant.io"

Restart Codex, confirm that the plugin skills remain available, and use /mcp or codex mcp list to verify that exactly one SocratiCode server is active. See bundled MCP server policy and the Codex configuration reference.

OpenAI Codex MCP-only installation

codex mcp add writes the user configuration in ~/.codex/config.toml:

codex mcp add socraticode -- npx -y --prefer-online socraticode@latest
codex mcp list

Start a new task or CLI session after installation. The equivalent TOML is:

[mcp_servers.socraticode]
command = "npx"
args = ["-y", "--prefer-online", "socraticode@latest"]

Both inline env = { ... } and a nested [mcp_servers.socraticode.env] table are valid. The CLI --env KEY=value option is usually clearer. See the OpenAI Codex MCP documentation.

VS Code Agent Plugin

This plugin bundles the MCP server, skills, and agent instructions for VS Code's native agent. It is separate from the SocratiCode editor extension.

1. Add this complete setting to the current VS Code profile before installation:

   json

   {
     "chat.plugins.enabled": true
   }
   

2. Run Chat: Install Plugin From Source from the Command Palette and enter https://github.com/giancarloerra/socraticode.

3. Start a new Chat session.

4. Verify SocratiCode under Agent Plugins - Installed, then run MCP: List Servers and confirm that its server is running.

Run Extensions: Check for Extension Updates to refresh installed agent plugins, then start a new Chat session. See VS Code Agent Plugins.

VS Code does not document a per-user environment overlay for a plugin-bundled MCP definition. To keep the plugin skills while applying custom variables, run MCP: List Servers, disable the bundled SocratiCode server, and add this user-scoped direct server through MCP: Open User Configuration:

{
  "servers": {
    "socraticode-configured": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "env": {
        "QDRANT_MODE": "external",
        "QDRANT_URL": "https://xyz.qdrant.io"
      }
    }
  }
}

Start a new Chat and use MCP: List Servers to confirm that only socraticode-configured is active. Use envFile instead of env when variables should come from a separate local file. Server enablement is stored separately from the shared plugin definition. See VS Code MCP server management and the env / envFile reference.

VS Code editor extension

The separately published editor extension adds the SocratiCode sidebar, status item, commands, walkthrough, and interactive graph webview. Install SocratiCode from the Visual Studio Marketplace in the current VS Code profile.

On Microsoft VS Code 1.99+ and compatible editors that implement the VS Code MCP provider API, the extension registers SocratiCode with the editor's native MCP registry. It does not configure independent clients such as Cline or Continue.

Enable VS Code's AI features and sign in to GitHub Copilot for native Agent Chat. On VS Code 1.122+, a BYOK model with tool calling also works without GitHub sign-in. Reload the window and start a new Chat session after installation. Run MCP: List Servers to confirm that SocratiCode is running, and open the SocratiCode sidebar to verify the editor UI. Update it through the Extensions view or Extensions: Check for Extension Updates.

The Open VSX package can be installed in VS Code-derived editors, but native MCP registration requires that editor to implement vscode.lm.registerMcpServerDefinitionProvider. See the VS Code MCP extension API.

VS Code direct MCP installation

Use the Stable or Insiders badge above, choose user or workspace scope in VS Code, then start a new Chat session. A project-scoped .vscode/mcp.json uses this complete object:

{
  "servers": {
    "socraticode": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"]
    }
  }
}

Verify with MCP: List Servers. Restart the server from that command and start a new Chat session after a release.

Cursor

The repository includes a Cursor-format plugin with the MCP server, skills, and instructions. SocratiCode is not currently published in the Cursor Marketplace, so use Cursor's documented user-scoped local-plugin directory. Clone the repository, then check out the stable version shown on the latest SocratiCode release:

mkdir -p ~/.cursor/plugins/local
git clone https://github.com/giancarloerra/socraticode.git ~/.cursor/plugins/local/socraticode
git -C ~/.cursor/plugins/local/socraticode checkout <latest-release-tag>

Replace <latest-release-tag> with the release tag shown on that page, for example vX.Y.Z. Restart Cursor or run Developer: Reload Window, then verify the plugin under Customize. To update to a later public release:

git -C ~/.cursor/plugins/local/socraticode fetch --tags
git -C ~/.cursor/plugins/local/socraticode checkout <latest-release-tag>

Reload Cursor after updating. These commands intentionally follow release tags rather than unreleased commits on main. See Cursor plugins.

For direct MCP configuration, use the Cursor badge above and select the intended user or project scope in Cursor. Start a new Agent chat, then verify socraticode under Customize → MCPs. The installation link already uses the latest-release engine command. See Cursor MCP install links.

The current local plugin does not declare user-configurable variables. To keep its skills while applying custom variables, open Customize, disable the plugin-provided SocratiCode MCP server, and add one direct server to the user or project mcp.json:

{
  "mcpServers": {
    "socraticode-configured": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "env": {
        "QDRANT_MODE": "external",
        "QDRANT_URL": "https://xyz.qdrant.io"
      }
    }
  }
}

Reload Cursor and verify under Customize that only socraticode-configured is active. See Cursor plugin variables and MCP server management.

The SocratiCode package on Open VSX is a VS Code-style editor extension, not a Cursor plugin. Installing that extension does not establish that Cursor implements VS Code's native MCP provider API. Use the local plugin or direct MCP path when MCP availability is required.

Gemini CLI

Add the user-scoped MCP server, verify it, then restart any active Gemini CLI session:

gemini mcp add --scope user socraticode npx -y --prefer-online socraticode@latest
gemini mcp list

For advanced configuration, edit the server in user scope (~/.gemini/settings.json) or define it in workspace scope (.gemini/settings.json) and explicitly forward the variables named in env:

{
  "mcpServers": {
    "socraticode": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "env": {
        "EMBEDDING_PROVIDER": "openai",
        "OPENAI_API_KEY": "${OPENAI_API_KEY}",
        "QDRANT_MODE": "external",
        "QDRANT_URL": "${QDRANT_URL}",
        "QDRANT_API_KEY": "${QDRANT_API_KEY}"
      }
    }
  }
}

Keep secret values in the process environment rather than committing them. Restart Gemini and run gemini mcp list to verify the configured server. See Gemini MCP configuration.

Continue

Continue uses a YAML mcpServers list. For project scope, create .continue/mcpServers/socraticode.yaml:

name: SocratiCode MCP
version: 1.0.0
schema: v1
mcpServers:
  - name: SocratiCode
    type: stdio
    command: npx
    args:
      - "-y"
      - "--prefer-online"
      - socraticode@latest

Continue refreshes saved configuration automatically. Open a new Continue Agent session and confirm the SocratiCode tools are listed. For user scope, add the same mcpServers list to ~/.continue/config.yaml. Continue can also import complete JSON MCP files placed in .continue/mcpServers/. See Continue MCP configuration and the Continue YAML reference.

Cline

For project scope, save this complete object as .cline/mcp.json. For user scope, add the same server to ~/.cline/data/settings/cline_mcp_settings.json through Cline's MCP settings interface:

{
  "mcpServers": {
    "socraticode": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "disabled": false,
      "autoApprove": []
    }
  }
}

Start a new Cline task and verify that socraticode and its tools appear in the MCP Servers view. Reconnect the server after a release. See the Cline MCP documentation.

Zed

Open Settings → AI → MCP Servers → Add Server → Add Local Server. The UI writes user-scoped settings. Use this complete server definition, either there or in project-scoped .zed/settings.json:

{
  "context_servers": {
    "socraticode": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "env": {}
    }
  }
}

Verify that the indicator beside SocratiCode is green and its tooltip says Server is active, then start a new Agent conversation. Restart the server from the MCP Servers page after a release.

Zed uses ~/.config/zed/AGENTS.md for personal instructions. For project instructions it uses the first matching supported file, which can be AGENTS.md, CLAUDE.md, GEMINI.md, or another supported compatibility file. Zed Rules were replaced by Skills and Instructions. See Zed MCP servers and Zed Instructions.

OpenCode

Use project-root opencode.json or opencode.jsonc for project scope. Use ~/.config/opencode/opencode.json or opencode.jsonc for user scope:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "socraticode": {
      "type": "local",
      "command": ["npx", "-y", "--prefer-online", "socraticode@latest"],
      "enabled": true
    }
  }
}

Restart OpenCode and verify the server with opencode mcp list. Restart OpenCode after a release so npm can check for the current engine. This is the OpenCode 1.x schema.

OpenCode V2 nests server names under mcp.servers and uses disabled instead of enabled; its local server definition is otherwise equivalent:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "servers": {
      "socraticode": {
        "type": "local",
        "command": ["npx", "-y", "--prefer-online", "socraticode@latest"],
        "disabled": false,
        "environment": {}
      }
    }
  }
}

See OpenCode 1.x MCP servers, OpenCode V2 MCP servers, and OpenCode configuration.

Other local stdio MCP hosts

The complete JSON object in Quick Start applies only to hosts whose documentation specifies an mcpServers object. Add it at the user or project scope documented by that host, restart the MCP server or start a new session, and verify SocratiCode in the host's MCP server and tool list. The example already checks for the latest published engine whenever the server starts. Hosts that support only remote HTTP MCP servers cannot launch SocratiCode directly.

Why SocratiCode

I built SocratiCode because I regularly work on existing, large, and complex codebases across different languages and need to quickly understand them and act. Existing solutions were either too limited, insufficiently tested for production use, or bloated with unnecessary complexity. I wanted a single focused tool that does deep codebase intelligence well — zero setup, no bloat, fully automatic — and gets out of the way.

Built-in Code Search vs SocratiCode

_{¹ VS Code Copilot: remote index via GitHub / Azure DevOps; local "External Ingest" gradually rolling out. ² LSP-based Find References / Go to Definition (Usages tool), not a full dependency graph. ³ Cursor: embeddings processed on Cursor servers (encrypted in transit and at rest). ⁴ VS Code Copilot: remote index hosted on GitHub / Azure DevOps. Sources: Cursor docs, Claude Code docs, VS Code Copilot docs.}

🔌 The context lives with your codebase, not with the assistant. Built-in indexes (Cursor's, Copilot's) are tied to that one tool — switch assistants and you start from scratch. SocratiCode is independent: index once, then plug it into Claude Code, Cursor, Copilot, Windsurf, your own private model, or all of them at once. They share the same understanding of your code.

On VS Code's 2.45M‑line codebase, SocratiCode answers architectural questions with 61% less data, 84% fewer steps, and 37× faster response than a grep‑based AI agent. Full benchmark →

Features

• Hybrid code search — Built on Qdrant, a purpose-built vector database with HNSW indexing, concurrent read/write, and payload filtering. Each chunk stores both a dense vector and a BM25 sparse vector; the Query API runs both sub-queries in a single round-trip and fuses results with Reciprocal Rank Fusion (RRF). Semantic search handles conceptual queries like "authentication middleware" even when those exact words don't appear in the code. BM25 handles exact identifier and keyword lookups. You get the best of both in every query with no tuning required.
• Configurable Qdrant — Use the built-in Docker Qdrant (default, zero config) or connect to your own instance (self-hosted, remote server, or Qdrant Cloud). Configure via QDRANT_MODE, QDRANT_URL, and QDRANT_API_KEY environment variables.
• Configurable Ollama — Use the built-in Docker Ollama (default, zero config) or point to your own Ollama instance (native install -GPU access-, remote server, etc.). Configure via OLLAMA_MODE, OLLAMA_URL, EMBEDDING_MODEL and EMBEDDING_DIMENSIONS environment variables.
• Multi-provider embeddings — Switch between Local Ollama (private, GPU access), Docker Ollama (zero-config), OpenAI (text-embedding-3-small, fastest), Google Gemini (gemini-embedding-001, free tier), LM Studio (local OpenAI-compatible server), or LiteLLM (proxy gateway in front of 100+ providers) with a single environment variable. No provider-specific configuration files.
• Private & secure — Everything runs on your machine — your code never leaves your network. The default Docker setup includes Ollama (embeddings) and Qdrant (vector storage) with no external API calls. No API costs, no token limits. Suitable for air-gapped and on-premises environments. Optional cloud providers (OpenAI, Google Gemini, Qdrant Cloud) are available but never required.
• AST-aware chunking — Files are split at function/class boundaries using AST parsing (ast-grep), not arbitrary line counts. This produces higher-quality search results. Falls back to line-based chunking for unsupported languages.
• Polyglot code dependency graph — Static analysis of import/require/use/include statements using ast-grep for 19+ languages. No external tools like dependency-cruiser required. Detects circular dependencies and generates visual Mermaid diagrams.
• Broad language support — Works with every supported file type out of the box. Fully supported languages bundle their grammars. GDScript uses an optional native parser when compatible and a syntax-aware fallback otherwise. For non-standard extensions, configure EXTRA_EXTENSIONS. If your AI can read it, SocratiCode can index it.
• Incremental indexing — After the first full index, only changed files are re-processed. Content hashes are persisted in Qdrant so state survives server restarts.
• Batched & resumable indexing — Files are processed in batches of 50, with progress checkpointed to Qdrant after each batch. If the process crashes or is interrupted, the next run automatically resumes from where it left off — already-indexed files are skipped via hash comparison. This keeps peak memory low and makes indexing reliable even for very large codebases.
• Live file watching — Optionally watch for file changes and keep the index updated in real time (debounced 2s). Watcher also invalidates the code graph cache.
• Parallel processing — Files are scanned and chunked in parallel batches (50 at a time) for fast I/O, while embedding generation and upserts are batched separately for optimal throughput.
• Multi-project — Index multiple projects simultaneously. Each gets its own isolated collection with full project path tracking.
• Cross-project search — Search across multiple related projects in a single query. Link projects via .socraticode.json or the SOCRATICODE_LINKED_PROJECTS env var, then set includeLinked: true on codebase_search. Results are tagged with project labels and ranked by cosine similarity, which is comparable across projects of very different sizes (falling back to rank fusion when a cosine is unavailable for any hit).
• Branch-aware indexing — Maintain separate indexes per git branch by setting SOCRATICODE_BRANCH_AWARE=true. Each branch gets its own Qdrant collections, so switching branches instantly switches to the correct index. Ideal for CI/CD pipelines and PR review workflows. Requires a path-derived project id: an explicit id (SOCRATICODE_PROJECT_ID, or projectId in .socraticode.json) is treated as a stable identity and is never suffixed, so branch-aware mode does not apply to those projects.
• Respects ignore rules — Honors all .gitignore files (root + nested), plus an optional .socraticodeignore for additional exclusions. Includes sensible built-in defaults. .gitignore processing can be disabled via RESPECT_GITIGNORE=false. Dot-directories (e.g. .agent) can be included via INCLUDE_DOT_FILES=true.
• Custom file extensions — Projects with non-standard extensions (e.g. .tpl, .blade) can be included via EXTRA_EXTENSIONS env var or extraExtensions tool parameter. Such files are indexed as plaintext and appear as leaf nodes in the code graph (no AST chunking or symbols). To instead treat a custom extension as a real language (full AST chunking, symbols, call graph), map it with EXTENSION_LANGUAGE_MAP (e.g. .inc:php).
• Configurable infrastructure — All ports, hosts, and API keys are configurable via environment variables. Qdrant API key support for enterprise deployments.
• Enterprise-ready simplicity — No agent coordination tuning, no memory limit environment variables, no coordinator/conductor capacity knobs, no backpressure configuration. SocratiCode scales by relying on production-grade infrastructure (Qdrant, proven embedding APIs) rather than complex in-process orchestration.
• Auto-setup & zero configuration — Just install the Claude Plugin/Skill or add the MCP server to your AI host config. On first use, the server automatically checks Docker, pulls images, starts Qdrant and Ollama containers, and downloads the embedding model. No config files, YAML, environment variables, or required native compilation. The optional GDScript parser falls back safely when no compatible native build is available. Works everywhere Docker runs.
• Session resume — By default, server startup resumes the indexed project represented by the MCP process's working directory. Complete indexes get a watcher plus an incremental catch-up update; interrupted indexes resume from the last checkpoint. Explicit project lists and SOCRATICODE_AUTO_RESUME=all extend this to other indexed projects.
• Auto-start watcher — In the default SOCRATICODE_WATCHER=auto mode, the file watcher starts during startup resume and after codebase_index or codebase_update. A completed indexed project not selected at startup gets a fallback watcher start on its first search, status, or graph interaction. manual permits only an explicit codebase_watch { action: "start" }; off disables watcher startup completely.
• Manual index snapshots — SOCRATICODE_WATCHER=off plus SOCRATICODE_AUTO_RESUME=off prevents implicit code-index updates, embeddings, and graph creation. Existing code indexes and graphs stay readable; refresh them explicitly with codebase_index, codebase_update, or codebase_graph_build.
• Git-triggered refresh: SOCRATICODE_WATCHER=git checks active indexed checkouts every 10 seconds and on search, status, and graph requests. A checked-out ref or HEAD change triggers the existing incremental updater and graph reconciliation. Native file watching stays disabled, including explicit watcher starts.
• Auto-build code graph — The code dependency graph is automatically built after indexing and rebuilt when watched files change. An update that changed nothing the graph is built from — a README, a fixture, a migration — reuses the existing graph rather than rebuilding it; each build records the inputs it read, and the rebuild is skipped only when every one of them is unchanged. Any graph-relevant addition rebuilds — a file appearing where the build looked, but not one the ignore rules exclude — and so does any configuration change, or a graph with no such record. No need to call codebase_graph_build manually unless you want to force a rebuild.
• Multi-agent collaboration — Multiple AI agents (each running their own MCP instance) can work on the same codebase simultaneously and share a single index. One agent triggers indexing, all agents search against the same data. Only one watcher runs per project — every agent benefits from real-time updates. Cross-process file locking coordinates indexing and watching automatically. Ideal for workflows like one agent writing tests while another fixes code, or a planning agent and an implementation agent working in parallel.
• Cross-process safety — File-based locking (proper-lockfile) prevents multiple MCP instances from simultaneously indexing or watching the same project. Stale locks from crashed processes are automatically reclaimed. When another MCP process is already watching a project, codebase_status reports "active (watched by another process)" instead of incorrectly showing "inactive."
• Concurrency guards — Duplicate indexing and graph-build operations are prevented. If you call codebase_index while indexing is already running, it returns the current progress instead of starting a second operation.
• Graceful stop — Long-running indexing operations can be stopped safely with codebase_stop. The current batch finishes and checkpoints, preserving all progress. Re-run codebase_index to resume from where it left off.
• Explicit reclamation — codebase_prune lists what the store holds and deletes one identity at a time, only on an explicit, confirmed apply. See Reclaiming stored identities.
• Graceful shutdown — On server shutdown, active indexing operations are given up to 60 seconds to complete, all file watchers are stopped cleanly, and the everything closes gracefully.
• Structured logging — All operations are logged with structured context for observability. Log level configurable via SOCRATICODE_LOG_LEVEL.
• Graceful degradation — If infrastructure goes down during watch, the watcher backs off and retries instead of crashing.

Prerequisites

Docker must be running when you use the server in the default managed mode.

The Qdrant container is managed automatically. If you set QDRANT_MODE=external and point QDRANT_URL at a remote or cloud Qdrant instance, Docker is only needed for Ollama (embeddings) in that case.

The Ollama container (embeddings) is also managed automatically in the default auto mode. SocratiCode first checks if Ollama is already running natively — if so it uses it. Otherwise it manages a Docker container for you. First-time download of the docker images or embedding models may take a few minutes, depending on your internet speed, and is required only at first launch.

Embedding performance on macOS / Windows

Docker containers on macOS and Windows cannot access the GPU (no Metal or CUDA passthrough). For small projects this is fine, but for medium-to-large codebases the CPU-only container is noticeably slower.

For best performance, install native Ollama: download and run the installer from ollama.com/download. Once Ollama is running, SocratiCode will automatically detect and use it — no extra configuration needed (first-time download of the embedding model, if not present, might take a few minutes). This gives you Metal GPU acceleration on macOS and CUDA on Windows/Linux.

If you prefer speed without a local install, see OpenAI Embeddings and Google Generative AI Embeddings below for cloud-based options. OpenAI is very fast with no local setup required. Google’s free tier is functional but rate-limited. See Environment Variables for configuration details.

Example Workflow

All tools default projectPath to the current working directory, so you never need to specify a path for the active project.

User: "Index this project"
→ codebase_index {}
  ⚡ Indexing started in the background — call codebase_status to check progress
→ codebase_status {}
  ⚠ Full index in progress — Phase: generating embeddings (batch 1/1)
  Progress: 247/1847 chunks embedded (13%) — Elapsed: 12s
→ codebase_status {}
  ✓ Indexing complete: 342 files, 1,847 chunks (took 115.2s)
  File watcher: active (auto-updating on changes)

User: "Search for how authentication is handled"
→ codebase_search { query: "authentication handling" }
  Runs dense semantic search + BM25 keyword search in parallel, fuses results with RRF
  Returns top 10 results ranked by combined relevance

User: "What files depend on the auth middleware?"
→ codebase_graph_query { filePath: "src/middleware/auth.ts" }
  Returns imports and dependents
  (graph was auto-built after indexing — no manual build needed)

User: "Show me the dependency graph"
→ codebase_graph_visualize {}
  Returns a Mermaid diagram colour-coded by language

User: "Are there any circular dependencies?"
→ codebase_graph_circular {}
  Found 2 cycles: src/a.ts → src/b.ts → src/a.ts

User: "What breaks if I rename validateUser?"
→ codebase_impact { target: "validateUser" }
  Blast radius for symbol: validateUser
  Hop 1 (3 files): src/auth/login.ts, src/api/users.ts, tests/auth.test.ts
  Hop 2 (5 files): ...

User: "What does the server entry point actually do?"
→ codebase_flow {}
  Detected 4 entry point(s):
    main (cmd/server.go:10) — well-known-name:main
    healthz (src/api/routes.ts:42) — framework:get
    ...
→ codebase_flow { entrypoint: "main" }
  └── main (cmd/server.go:10)
      ├── loadConfig (cmd/server.go:15)
      └── startServer (src/server.ts:8)
          └── ...

User: "Who calls bcryptCompare and what does it call?"
→ codebase_symbol { name: "bcryptCompare" }
  Symbol: bcryptCompare (function)
  Defined: src/auth/hash.ts:42–58
  Callers (3): ← src/auth/login.ts:12, ← src/auth/reset.ts:30 ...
  Callees (1): → compare [unique, 1 candidate]

Agent Instructions

Claude Code plugin users: These instructions are included automatically as skills in the SocratiCode plugin. You don't need to copy them into CLAUDE.md. The section below is for non-Claude Code hosts (VS Code, Cursor, Claude Desktop, etc.).

For best results, add instructions like the following to your AI assistant's project-level instructions file. The core principle: search before reading. The index gives you a map of the codebase in milliseconds; raw file reading is expensive and context-consuming.

Where to place these instructions (per IDE):

Why this matters: Installing the MCP server alone gives your agent access to SocratiCode tools, but the agent still decides when to use them. Adding these instructions to your project ensures the agent consistently prefers SocratiCode search over raw file reads, uses the graph for dependency-aware tasks, and follows the search-before-reading workflow.

## Codebase Search (SocratiCode)

This project is indexed with SocratiCode. Always use its MCP tools to explore the codebase
before reading any files directly.

### Workflow

1. **Start most explorations with `codebase_search`.**
   Hybrid semantic + keyword search (vector + BM25, RRF-fused) runs in a single call.
   - Use broad, conceptual queries for orientation: "how is authentication handled",
     "database connection setup", "error handling patterns".
   - Use precise queries for symbol lookups: exact function names, constants, type names.
   - Prefer search results to infer which files to read — do not speculatively open files.
   - **When to use grep instead**: If you already know the exact identifier, error string,
     or regex pattern, grep/ripgrep is faster and more precise — no semantic gap to bridge.
     Use `codebase_search` when you're exploring, asking conceptual questions, or don't
     know which files to look in.

2. **Follow the graph before following imports.**
   Use `codebase_graph_query` to see what a file imports and what depends on it before
   diving into its contents. This prevents unnecessary reading of transitive dependencies.
   - **Before modifying or deleting a file**, check its dependents with `codebase_graph_query`
     to understand the blast radius.
   - **When planning a refactor**, use the graph to identify all affected files before
     making changes.

3. **Use Impact Analysis BEFORE refactoring, renaming, or deleting code.**
   The symbol-level call graph (`codebase_impact`, `codebase_flow`, `codebase_symbol`,
   `codebase_symbols`) goes one step deeper than the file graph: it knows which
   functions and methods call which.
   - `codebase_impact` answers "what breaks if I change X?" (blast radius — every file
     that transitively calls into the target).
   - `codebase_flow` answers "what does this code do?" by tracing forward from an entry
     point. Call with no `entrypoint` to discover candidate entry points (auto-detected
     via orphans, conventional names like `main()`, framework routes, tests).
   - `codebase_symbol` gives a 360° view of one function: definition, callers, callees.
   - `codebase_symbols` lists symbols in a file or searches by name.
   - Always prefer these over reading multiple files when the question is about
     dependencies between functions, not concepts.

4. **Read files only after narrowing down via search.**
   Once search results clearly point to 1–3 files, read only the relevant sections.
   Never read a file just to find out if it's relevant — search first.

5. **Use `codebase_graph_circular` when debugging unexpected behaviour.**
   Circular dependencies cause subtle runtime issues; check for them proactively.
   Also run `codebase_graph_circular` when you notice import-related errors or unexpected
   initialisation order.

6. **Check `codebase_status` if search returns no results.**
   The project may not be indexed yet. Run `codebase_index` if needed, then wait for
   `codebase_status` to confirm completion before searching.

7. **Leverage context artifacts for non-code knowledge.**
   Projects can define a `.socraticodecontextartifacts.json` config to expose database
   schemas, API specs, infrastructure configs, architecture docs, and other project
   knowledge that lives outside source code. These artifacts are auto-indexed alongside
   code during `codebase_index` and `codebase_update`.
   - Run `codebase_context` early to see what artifacts are available.
   - Use `codebase_context_search` to find specific schemas, endpoints, or configs
     before asking about database structure or API contracts.
   - If `codebase_status` shows artifacts are stale, run `codebase_context_index` to
     refresh them.

### When to use each tool

| Goal | Tool |
|------|------|
| Understand what a codebase does / where a feature lives | `codebase_search` (broad query) |
| Find a specific function, constant, or type | `codebase_search` (exact name) or grep if you know already the exact string |
| Find exact error messages, log strings, or regex patterns | grep / ripgrep |
| See what a file imports or what depends on it | `codebase_graph_query` |
| Check blast radius before modifying or deleting a file | `codebase_impact` (symbol-level) or `codebase_graph_query` (file-level) |
| **What breaks if I change function X?** | `codebase_impact target=X` |
| **What does this entry point actually do?** | `codebase_flow entrypoint=X` |
| **List entry points in this codebase** | `codebase_flow` (no args) |
| **Who calls this function and what does it call?** | `codebase_symbol name=X` |
| **What functions/classes exist in this file?** | `codebase_symbols file=path` |
| **Search for symbols by name across the project** | `codebase_symbols query=X` |
| Spot architectural problems | `codebase_graph_circular`, `codebase_graph_stats` |
| Visualise module structure | `codebase_graph_visualize` |
| Verify index is up to date | `codebase_status` |
| Discover what project knowledge (schemas, specs, configs) is available | `codebase_context` |
| Find database tables, API endpoints, infra configs | `codebase_context_search` |

Why semantic search first? A single codebase_search call returns ranked, deduplicated snippets from across the entire codebase in milliseconds. This gives you a broad map at negligible token cost — far cheaper than opening files speculatively. Once you know which files matter, targeted reading is both faster and more accurate. That said, grep remains the right tool when you have an exact string or pattern — use whichever fits the query.

Keep the connection alive during indexing. Indexing runs in the background — the MCP server continues working even when not actively responding to tool calls. However, some MCP hosts might disconnect an idle MCP connection after a period of inactivity, which might cut off the background process. Instruct your AI to call codebase_status roughly every 60 seconds after starting codebase_index until it completes. This keeps the host connection active and provides real-time progress.

Configuration

Install

Use the host-specific steps in Plugins and host integrations. They cover installation scope, activation, verification, updates, and each host's actual configuration schema.

From source (for contributors)

git clone https://github.com/giancarloerra/socraticode.git
cd socraticode
npm install
npm run build

Register node /absolute/path/to/socraticode/dist/index.js in the user or project scope supported by your MCP host, then restart the server or start a new session. Verify SocratiCode in the host's MCP server list. To update, run git pull --ff-only, npm install, and npm run build in the clone, then restart the MCP server and verify it again.

MCP host config variants

The examples below use the conventional JSON mcpServers shape to show SocratiCode settings. Apply the same command and environment values through the host-specific schema documented in Plugins and host integrations. Continue, Gemini CLI, VS Code, Zed, and OpenCode use different configuration paths or wrappers.

Default (zero config, from source)

Using npx? Replace the node command and source path below with "command": "npx" and "args": ["-y", "--prefer-online", "socraticode@latest"].

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"]
    }
  }
}

Tip: The default OLLAMA_MODE=auto detects native Ollama (port 11434) on startup and uses it if available, otherwise falls back to a managed Docker container. To make your config self-documenting, add an "env" block with explicit values. See Environment Variables for all options.

External Ollama (native install)

If you have Ollama installed natively, set OLLAMA_MODE=external and point to your instance:

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "OLLAMA_MODE": "external",
        "OLLAMA_URL": "http://localhost:11434"
      }
    }
  }
}

The embedding model is pulled automatically on first use. To pre-download: ollama pull nomic-embed-text

Remote Ollama server

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "OLLAMA_MODE": "external",
        "OLLAMA_URL": "http://gpu-server.local:11434"
      }
    }
  }
}

OpenAI Embeddings

Use OpenAI's cloud embedding API instead of local Ollama. Requires an API key.

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "EMBEDDING_PROVIDER": "openai",
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Defaults: EMBEDDING_MODEL=text-embedding-3-small, EMBEDDING_DIMENSIONS=1536. For higher quality, use text-embedding-3-large with EMBEDDING_DIMENSIONS=3072.

Google Generative AI Embeddings

Use Google's Gemini embedding API. Requires an API key.

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "EMBEDDING_PROVIDER": "google",
        "GOOGLE_API_KEY": "AIza..."
      }
    }
  }
}

Defaults: EMBEDDING_MODEL=gemini-embedding-001, EMBEDDING_DIMENSIONS=3072.

LM Studio (local, OpenAI-compatible)

LM Studio ships with a Local Server that exposes an OpenAI-compatible API on http://localhost:1234/v1. Use this provider when you want to host embedding models in LM Studio (e.g. when LM Studio is your single source for both chat and embedding models, or when you want a Mac/Windows-friendly desktop UI for managing GGUF models).

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "EMBEDDING_PROVIDER": "lmstudio",
        "EMBEDDING_MODEL": "nomic-embed-text-v1.5",
        "EMBEDDING_DIMENSIONS": "768"
      }
    }
  }
}

No defaults — EMBEDDING_MODEL and EMBEDDING_DIMENSIONS are required. LM Studio has no out-of-the-box embedding model; you load one yourself in the Local Server tab. SocratiCode fails fast if either is missing.

Optional: LMSTUDIO_URL (default http://localhost:1234/v1) for non-default ports; LMSTUDIO_API_KEY if you've enabled API key auth in LM Studio; LMSTUDIO_ALLOW_MISSING_MODEL_LISTING=true for OpenAI-compatible servers that have no /v1/models endpoint (see below).

This provider also drives any other server that speaks the OpenAI embeddings API. Single-model servers such as HuggingFace Text Embeddings Inference (TEI) fix the model at startup and answer /v1/models with a 404, so readiness needs LMSTUDIO_ALLOW_MISSING_MODEL_LISTING=true to fall back to probing /v1/embeddings:

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "EMBEDDING_PROVIDER": "lmstudio",
        "LMSTUDIO_URL": "http://localhost:8080/v1",
        "EMBEDDING_MODEL": "BAAI/bge-m3",
        "EMBEDDING_DIMENSIONS": "1024",
        "LMSTUDIO_ALLOW_MISSING_MODEL_LISTING": "true"
      }
    }
  }
}

EMBEDDING_MODEL is whatever the server was started with (TEI's --model-id) and EMBEDDING_DIMENSIONS must match that model's output width — the probe checks it and fails fast on a mismatch, since without /v1/models there is nothing else to verify against.

LiteLLM (proxy gateway, 100+ providers)

LiteLLM Proxy Server exposes an OpenAI-compatible /v1/embeddings endpoint and fans out to any of 100+ underlying providers (OpenAI, Anthropic, Cohere, Voyage, HuggingFace, Bedrock, Vertex AI, Ollama, ...). Use this provider when you want centralised key management (one virtual key per developer instead of N provider keys spread across MCP configs), fallback / load balancing between embedding backends, or provider-agnostic indexes that survive a backend swap.

{
  "mcpServers": {
    "socraticode": {
      "command": "node",
      "args": ["/absolute/path/to/socraticode/dist/index.js"],
      "env": {
        "EMBEDDING_PROVIDER": "litellm",
        "LITELLM_API_KEY": "sk-...",
        "EMBEDDING_MODEL": "text-embedding-3-small",
        "EMBEDDING_DIMENSIONS": "1536"
      }
    }
  }
}

LITELLM_API_KEY, EMBEDDING_MODEL, and EMBEDDING_DIMENSIONS are all required. LiteLLM proxies always authenticate (master key or virtual key from /key/generate); the alias name and underlying dimension come from your config.yaml. SocratiCode fails fast on any missing piece.

Optional: LITELLM_URL (default http://localhost:4000/v1) — must include the /v1 suffix; LITELLM_SEND_DIMENSIONS=true to forward the OpenAI dimensions parameter through the proxy (only safe for Matryoshka-aware backends like text-embedding-3-* or voyage-3 — non-Matryoshka backends reject the request).

This is a client for the LiteLLM proxy server, not the LiteLLM Python library, and it does not route provider/model strings itself. It sends EMBEDDING_MODEL to LITELLM_URL verbatim and requires that name to appear in the proxy's /v1/models. To reach a backend such as OpenRouter, register it in the proxy's config.yaml model_list (set model_name to the value you put in EMBEDDING_MODEL, and litellm_params.model to e.g. openrouter/qwen/qwen3-embedding-8b); the proxy does the routing and SocratiCode just sends the alias. Pointing LITELLM_URL directly at a non-LiteLLM endpoint works only if that endpoint is OpenAI-compatible, lists your EMBEDDING_MODEL under /v1/models, and accepts it under its own native model id (no LiteLLM provider/ prefix).

Git Worktrees (shared index across directories)

If you use git worktrees — or any workflow where the same repository lives in multiple directories — each path would normally get its own Qdrant index. This means redundant embedding and storage for what is essentially the same codebase.

Set SOCRATICODE_PROJECT_ID to share a single index across all directories of the same project.

MCP hosts with git worktree detection (e.g. Claude Code)

Some MCP hosts (like Claude Code) resolve the project root by following git worktree links. Since worktrees point back to the main repository's .git directory, the host automatically maps all worktrees to the same project config. This means you only need to configure the MCP server once for the main checkout — all worktrees inherit it automatically.

For Claude Code, add the server with local scope from your main checkout:

cd /path/to/main-checkout
claude mcp add -e SOCRATICODE_PROJECT_ID=my-project --scope local socraticode -- npx -y --prefer-online socraticode@latest

All worktrees created from this repo will automatically connect to socraticode with the shared project ID. No per-worktree setup needed.

Note: This only works for git worktrees. Separate git clones of the same repo have independent .git directories and won't share the config.

Other MCP hosts (per-project .mcp.json)

For MCP hosts that don't resolve git worktree paths, add a .mcp.json at the root of each worktree (and your main checkout):

{
  "mcpServers": {
    "socraticode": {
      "command": "npx",
      "args": ["-y", "--prefer-online", "socraticode@latest"],
      "env": {
        "SOCRATICODE_PROJECT_ID": "my-project"
      }
    }
  }
}

Add .mcp.json to your .gitignore if you don't want it tracked.

How it works

With this config, agents running in /repo/main, /repo/worktree-feat-a, and /repo/worktree-fix-b all share the same codebase_my-project, codegraph_my-project, and context_my-project Qdrant collections.

How it works in practice:

• The semantic index reflects whichever worktree last triggered a file change — but since branches typically differ by only a handful of files, the index is 99%+ accurate for all worktrees
• Your AI agent reads actual file contents from its own worktree; the shared index is only used for discovery and navigation
• When changes merge back to main, the file watcher re-indexes the changed files and the index converges

Team-Shared Index (committed projectId)

The env-var approach above works per-machine. For a stable identifier that every teammate (and CI runner) picks up automatically, commit a projectId in .socraticode.json at the project root:

{
  "projectId": "my-project"
}

Now any checkout of the repo — regardless of where it lives on disk or which user account owns it — addresses the same codebase_my-project, codegraph_my-project, and context_my-project Qdrant collections. This is the recommended setup for teams sharing a Qdrant instance: the index is built once and benefits everyone, even across different OS users and laptops with completely different filesystem layouts.

The value must match [a-zA-Z0-9_-]+; whitespace is trimmed, and a missing or empty value falls back to the path-hash default. The SOCRATICODE_PROJECT_ID env var, when set, takes precedence over this file — handy for ad-hoc per-machine overrides without touching the repo.

Cross-Project Search (linked projects)

If you work across multiple related repositories or packages, you can search them all in a single query.

Configuration

Create a .socraticode.json file in your project root:

{
  "linkedProjects": [
    "../shared-lib",
    "/absolute/path/to/other-project"
  ]
}

Or set the SOCRATICODE_LINKED_PROJECTS environment variable (comma-separated paths):

SOCRATICODE_LINKED_PROJECTS="../shared-lib,/absolute/path/to/other-project"

Both sources are merged and deduplicated. Relative paths are resolved from the project root. Non-existent paths are silently skipped.

Usage

Pass includeLinked: true to codebase_search:

Search for "authentication middleware" with includeLinked: true

Results are ranked by cosine similarity against the query, not by each hit's position within its own project. A rank only means something inside the list it came from, so ranking on it let the top hit of a small project outrank a far stronger hit from a large one, and capped every cross-project score at 1/61 ≈ 0.016 — below the SEARCH_MIN_SCORE default of 0.10, which then discarded everything. Cosine is an absolute measure against the same query vector, so it is comparable across projects and lands on the same scale the threshold expects. Single-project search is unaffected and still uses Qdrant's server-side RRF.

If a cosine cannot be computed for every hit, ranking falls back to the previous rank fusion for that whole query rather than mixing two different measures. That is detected in three cases, each logged with the collection involved: a chunk returned without a usable dense vector, a vector of zero magnitude, and a vector whose dimensionality differs from the query's. Note the limit of that last one: it catches a collection embedded with a model of a different dimensionality, but a different model producing the same dimensionality is indistinguishable here and would still be scored, so keep linked projects on one embedding model.

In the fallback, a single hit contributes at most 1/(60+0+1) ≈ 0.0164, though a file matching as several chunks accumulates one contribution per chunk (1/61 + 1/62 ≈ 0.0325), so final scores can exceed that bound. Either way they sit far below the 0.10 default, so lower minScore for that query if you hit it.

Results are tagged with [project-name] labels showing which project each result came from. Deduplication is scoped to a single project: the same relative path in two different projects is kept as two separate results, because they are genuinely different files (your own src/util.ts and a linked project's are not interchangeable). Within one project, the higher-priority occurrence of a path wins.

Note: Each linked project must be independently indexed (codebase_index) before it can be searched.

Python Import Roots

Python projects whose import root is not the project root, src/, lib/, or a root declared by pyproject.toml can list their runtime path entries in the project-root .socraticode.json. For example, Apache Airflow adds dags/ to sys.path, so a DAG repository can declare:

{
  "pythonRoots": ["dags"]
}

Entries must name existing project-root-relative directories. They are trimmed, normalized, and deduplicated while preserving their declared order. Absolute paths, paths outside the project root (including escaping symlinks), missing paths, files, and non-string values are ignored. If pythonRoots is absent, empty, or has no usable entry, Python resolution behaves exactly as before.

Resolution keeps the existing probes first: the project root, project-root src/ and lib/, and the importing file's directory. Configured roots are then tried in their declared order, followed by roots derived from applicable pyproject.toml manifests. This lets an explicit runtime layout take precedence over manifest-derived alternatives without changing existing projects.

Adding, removing, reordering, or changing the effective roots rebuilds the code graph on the next incremental update. It does not rebuild or invalidate the code/vector index, and no manual re-index is required.

Branch-Aware Indexing

By default, all branches of a project share the same index. When you switch branches, changed files are re-indexed by the watcher, and the index reflects the current branch state.

For workflows where you need separate, persistent indexes per branch — such as CI/CD pipelines or comparing code across branches — enable branch-aware mode:

SOCRATICODE_BRANCH_AWARE=true

With this enabled, collection names include the branch name (e.g. codebase_abc123__main, codebase_abc123__feat_my-feature). Each branch maintains its own independent index, code graph, and context artifacts.

Also requires a usable branch name. Branch names are sanitized for use in collection names, so a branch made only of separators (e.g. ___) reduces to an empty suffix and no per-branch collection is created. Where that happens, set a branch-specific SOCRATICODE_PROJECT_ID instead.

Only applies to path-derived project ids. If the project pins an id — via SOCRATICODE_PROJECT_ID or projectId in .socraticode.json — that id is treated as a stable identity and is never given a branch suffix, so every branch continues to share one index. This is deliberate: suffixing an explicit id would change the collection names an existing installation already uses and make its indexes appear missing. SocratiCode logs a warning once per project when branch-aware mode is enabled but ignored for this reason. To get per-branch collections, remove the explicit id so the path-derived id can carry the suffix, or set a branch-specific SOCRATICODE_PROJECT_ID deliberately.

When to use:

• CI/CD pipelines that index each branch/PR separately
• Comparing search results across branches
• Keeping a pristine main index unaffected by feature branch changes

When NOT to use:

• Local development with frequent branch switching (default shared index is more efficient)
• Projects tracked via SOCRATICODE_PROJECT_ID (explicit IDs bypass branch detection)

How it works: projectIdFromPath() detects the current git branch via git rev-parse --abbrev-ref HEAD and appends a sanitized branch suffix (e.g. feat/my-feature → feat_my-feature) to the hash-based project ID. Detached HEAD states fall back to the branchless ID.

Reclaiming stored identities

An index outlives its directory when a worktree is removed, a clone deleted or a folder renamed: codebase_remove addresses a project by its path, so a directory that is gone can no longer be named. codebase_prune works from what the store already records.

Report. codebase_prune with no arguments is a neutral inventory. For every stored identity it lists the collections that carry its name (code, graph, symbol graph, context) and every metadata record, and it reports the recorded path as one of:

Two identities recording the same canonical path — a path-hash identity and a later pinned projectId for the same checkout — are marked possible-superseded as an advisory only. An identity whose records disagree about its path is marked for manual inspection and cannot be deleted by this tool. Metadata points that cannot be attributed to an identity are listed with their point id and the reason, and are likewise left alone. So is a collection whose name fits two identities: a pinned projectId that itself begins with codebase_, codegraph_ or context_, or ends with _symgraph_meta/_file/_index, produces names that another identity could also own; the inventory settles them with what the store shows (a metadata point names a family collection, a full symbol-graph triple names a symbol graph) and lists the rest for manual inspection; every identity such a name could belong to is held back from deletion as well, since deleting one would remove the evidence and hand the collection to the other on the next read.

Apply. Deletion takes the exact identity from the report, the confirmationToken printed beside it, and acknowledgeNoRemoteWriters: true. The token fingerprints every collection and every metadata record of that identity, so any change since the report — a new record, a changed status, a resource added or gone — makes the apply refuse and ask for a fresh report. Deletion is refused while the identity is being indexed, watched, graph-built or context-indexed in this process, while another process on this host holds any of its writer locks (index, watch, graph, context), or whenever a lock cannot be inspected. From the final validation through the last delete the tool holds an identity-scoped barrier: the prune lock plus every writer lock. Every writer checks the barrier before starting and takes its own lock non-reentrantly while it runs (a lock this process already holds is never adopted), so a graph build or context indexing in another local process is excluded the same way an indexer or watcher is; a second process asking to build the same identity's graph while one is running is refused rather than run alongside it.

Shared-store limitation. Lock files live on this host. They cannot show that another machine sharing the same Qdrant is idle, which is what acknowledgeNoRemoteWriters states on the operator's behalf; nothing in the tool infers it.

Outcome. Every resource is reported with its own result. Barrier ownership is checked before each write; if a barrier lock is reported lost mid-cleanup, deletion stops at the next write, the remaining resources are reported as skipped, and the cleanup is reported incomplete. A resource already gone counts as deleted, so repeating an apply after a partial failure is safe and does nothing where nothing remains. Success is reported only when the inventory re-read after deletion shows none of the identity's resources; otherwise the leftovers are listed. Caches keyed by the identity — the graph, the loaded file hashes and the symbol graph — are dropped in this process. Existing indexes need no migration or re-index to be reported.

Optional automatic local cleanup

Automatic cleanup is disabled by default. Set SOCRATICODE_AUTO_CLEANUP=local only when Qdrant is host-exclusive and not shared with another machine. This explicitly declares local ownership; it requires managed mode and a loopback endpoint. Shared/external stores and explicit project IDs remain operator-controlled through codebase_prune.

New eligible indexes record a local installation UUID plus their original repository, exact branch ref, checkout and Git administration directory. The UUID is stored in .socraticode/local-index-owner under the user's home directory and reused on restart. Existing records and orphan resources without sufficient proof remain automatic-report-only, with no migration or re-index required. Losing the local ownership file does not authorize adopting those records. Index metadata written while cleanup is disabled does not acquire automatic-cleanup ownership.

The MCP server checks at startup and every minute, independently of auto-resume. It reuses prune's inventory fingerprint, writer barriers, resource deletion and verification. A linked worktree qualifies only after Git unregisters it, its original Git administration directory is removed and its source directory is absent on the same verified accessible filesystem. A branch-aware index qualifies only after its exact local ref is deleted and no registered worktree uses it. A merge, squash merge, branch switch, moved worktree, missing directory alone, unavailable mount, permission error, ambiguous identity or active writer never authorizes deletion. Only index resources and metadata are removed, never source directories or Git branches. Deletions, refusals and incomplete cleanup are reported through MCP log notifications. Partial collection failures retain ownership metadata for a safe subsequent retry.

Available tools

Once connected, the following tools are available to your AI assistant:

Indexing

Search

Code Graph

Impact Analysis (symbol-level call graph)

A second graph layer goes one step deeper than file imports — it tracks which functions and methods call which. Use these tools BEFORE refactoring, renaming, or deleting code.

Accepted limits. The call graph is static-analysis-based — no type inference. Dynamic dispatch (getattr, obj[key](...), reflection, eval), unexpanded macros, and framework magic (Spring @Autowired, Angular DI, Rails has_many, decorator-driven routing) are invisible. Callers that reach a method only through these mechanisms will not appear in codebase_impact. Treat "zero callers" as a hint to double-check on DI-heavy codebases. codebase_graph_status reports the unresolved share (unresolvedEdgePct) as a quality signal: the percentage of captured symbol edges (calls, imports, re-exports and type or value references) that matched no project symbol, GDScript engine calls excluded. That share includes edges into runtime builtins and external libraries (setTimeout, JSON.stringify, an SDK client's methods), which no project symbol can resolve, so it runs high on healthy TypeScript/JavaScript code and is not a resolver failure rate. See DEVELOPER.md § Impact Analysis for the full list.

Interactive graph explorer

Ask your AI "show me an interactive graph of this project" (or invoke codebase_graph_visualize with mode: "interactive") and SocratiCode generates a self-contained HTML page and opens it in your default browser:

• File view — every source file as a node, imports as edges, language-coloured, circular deps in red.
• Symbol view — toggle to see functions/classes/methods as nodes with call edges (available when the symbol graph fits within the embed cap; above that threshold the file view remains and the banner points at codebase_impact for symbol-level queries).
• Sidebar — click a node to see imports / dependents / symbols-in-file / line numbers, with action buttons for blast radius and call flow.
• Right-click any node → highlights its reverse-transitive closure (who breaks if this changes).
• Live search filters and centres matching nodes. Layout switcher — Dagre / force-directed / concentric / breadth-first / grid / circle. Export PNG produces a shareable image.
• Offline-safe — Cytoscape.js + Dagre are vendored inside the SocratiCode package. No CDN, no network, works in air-gapped environments.

The output is a single HTML file (written to the OS temp dir, one per project) that you can also commit to a PR or share on Slack.

Management

Context Artifacts

Language Support

SocratiCode supports languages at three levels:

Full Support (indexing + code graph + AST chunking)

JavaScript, TypeScript, TSX, Python, Java, Kotlin, Scala, C, C++, C#, Go, Rust, Ruby, PHP, Swift, Dart, Elixir (including HEEx/EEx), Bash/Shell, HTML, CSS/SCSS, Svelte, Vue

Conditional AST Support

GDScript: .gd files use an optional native binary (tree-sitter-gdscript) for AST chunking, symbol extraction, and call-site resolution. When the native parser is unavailable, a syntax-aware fallback handles import extraction and line-based chunking without creating edges from comments or strings. class_name declarations, extends/preload/load imports, and res:// paths are resolved. Relative extends and preload paths use the script directory; relative runtime load paths use the Godot project root. Per-file project.godot discovery supports nested and sibling Godot projects. Godot scenes and resources (.tscn/.tres) are indexed with line-based chunking. [ext_resource] declarations are extracted as dependency edges via a tokenizer that handles arbitrary whitespace, attribute order, escaped quotes, and both uid="uid://..." and path="..." attributes. When both are present, the UID takes priority. Both res:// paths and paths relative to the .tscn/.tres file are supported, per the TSCN documentation. res:// paths resolve to .gd, .tscn, and .tres targets in the code graph.

Svelte and Vue: imports extracted from <script> blocks (re-parsed as TypeScript) and CSS @import/@require from <style> blocks (any combination of lang, scoped, module, global attributes). Path aliases from tsconfig.json/jsconfig.json compilerOptions.paths are resolved (including extends chains). SCSS partial resolution (_ prefix convention) is supported.

Python: absolute imports resolve through the import roots implied by the project's pyproject.toml files — root and nested, so uv workspaces get cross-package edges — covering both the src/ layout (packages/<dist>/src/<module>/…, what uv init --lib, hatchling and setuptools generate) and the flat layout beside each manifest, including PEP 420 namespace packages and single-module distributions. A manifest applies to a file only when it sits on that file's ancestor path or an ancestor manifest declares it a [tool.uv.workspace] member, so a sample app, docs project or checked-in sdist carrying its own manifest does not become an import root for unrelated code; only [tool.uv.workspace] is read, so poetry, pdm and hatch path-dependency monorepos get ancestor-path scoping and no cross-package edges; applicable roots are tried nearest first, so a package resolves its own modules before a sibling's. Relative imports and the project-root src//lib/ and sibling-flat conventions are unchanged — the sibling-flat fallback still takes precedence over these roots, matching CPython, which puts the script's own directory at sys.path[0].

PHP: use imports resolve through the PSR-4 prefixes declared in the project's composer.json files — root and nested, so a Composer monorepo's path packages get cross-package edges. Where no prefix matches, they resolve against the namespace and class/interface/trait/enum declarations found in the project itself, which is what reaches a package that ships "autoload": {} and registers its namespaces at run time ($loader->addNamespace(...), the WordPress-plugin norm) — no spl_autoload_register interpretation involved. Comma lists (use A\B, A\C;), groups (use A\{B, C};, including per-member function/const modifiers) and fully-qualified names (use \A\B;) are all read. require/include resolve relative paths, bare paths (source directory first, then the project root), and __DIR__ . '<literal>' / dirname(__FILE__) . '<literal>', the dominant include idiom outside Composer. They are read from the include expressions themselves, so they are found in any position — return require __DIR__ . '/routes.php'; and $config = include 'config.php'; count, while an include mentioned in a comment or quoted inside a string does not. An include joined to a run-time value (ABSPATH . '/x.php', $base . '/x.php') stays unresolved rather than guessed.

Dart: symbols (classes, mixins, enums, extensions, typedefs, functions, getters, setters, operators, constructors including named and factory, and abstract/bodyless members), call sites (method calls, cascades, constructor invocations), main() entry-point detection, and AST chunking are all tree-sitter based; import/export/part edges are extracted via regex. Intra-project package: imports (the Flutter convention) resolve through the project's pubspec.yaml files — root and nested, so pub-workspace/melos monorepos get cross-package edges — via pub's package:<name>/<rest> → <package_root>/lib/<rest> mapping; dart: and unknown package names stay external. The bundled grammar (@ast-grep/lang-dart) predates Dart 3 class modifiers (sealed/base/interface/final/mixin class) and extension type: declarations using those are skipped (with a one-time warning logged) until the upstream grammar is updated, while the rest of each file still indexes normally.

Elixir: .ex and .exs files use the ast-grep grammar for chunking. alias, import, require, and use directives resolve to in-project defmodule declarations; fully qualified calls without one of those directives remain unresolved. defmodule, def, defp, and ordinary calls produce symbols and call edges. Chunking is module-level, with large modules falling back to line windows rather than per-function chunks. defprotocol, defimpl, defguard, defmacro, and defdelegate declarations do not register symbols yet; protocol and implementation scopes are not represented, so functions inside them appear as top-level symbols. Standalone .heex and .eex templates use dedicated tree-sitter grammars for AST chunking, remote-component dependencies, and calls from embedded Elixir expressions; markup and comments are never parsed as Elixir. .leex uses the EEx grammar on a best-effort basis, safely falling back to line chunks and no extracted edges when parsing fails.

Code Graph via Regex + Indexing

Lua (require/dofile/loadfile), SASS, LESS, Stylus (CSS @import/@require extraction)

Indexing Only (hybrid search, line-based chunking)

JSON, YAML, TOML, XML, INI/CFG, Markdown/MDX, RST, SQL, R, Dockerfile, TXT, and any file matching a supported extension or special filename (Dockerfile, Makefile, Gemfile, Rakefile, etc.)

63 file extensions + 8 special filenames supported out of the box.

Extensionless files (Unix scripts, health probes, sourced libraries) are also indexed via content-based language detection when INDEX_EXTENSIONLESS is enabled (the default) — see that environment variable below.

Ignore Rules

The indexer combines three layers of ignore rules:

1. Built-in defaults — node_modules, .git, dist, build, lock files, IDE folders, etc.
2. .gitignore — All .gitignore files in the project (root and nested subdirectories). Set RESPECT_GITIGNORE=false to skip .gitignore processing entirely.
3. .socraticodeignore — Optional file for indexer-specific exclusions. Same syntax as .gitignore.

All three layers also apply to a context artifact that points at a directory, but they are resolved relative to the artifact directory, not the project root. A project-root .gitignore or .socraticodeignore governs the code index and does not reach a directory artifact. What applies to a directory artifact is the built-in defaults, the .gitignore at the artifact root and any nested .gitignore files, and a .socraticodeignore only at the artifact root. Nested .socraticodeignore files are not read. To exclude something from a directory artifact, put the pattern in one of those applicable files.

Note that a directory artifact inherits the built-in defaults in full, not just the build-output ones. Beyond __pycache__, *.pyc, dist and build, that list also covers names an artifact directory might legitimately use: env, vendor, target, out, coverage, *.map, *.log. If a directory artifact needs one of those, re-include it with a ! pattern in the .socraticodeignore at the artifact root, negating the name itself (!env). Gitignore semantics cannot re-include a file whose parent directory is excluded, so !env/** on its own does nothing. target has one additional constraint: its contents can be re-included, but target/ is skipped while discovering nested .gitignore files, so rules from target/.gitignore are not loaded. Put those rules in the .gitignore or .socraticodeignore at the artifact root instead. Files dropped by the ignore rules, by the binary check, or because they could not be read are counted in that artifact's log line when it is indexed. node_modules, .git and dot-files are pruned before the walk sees them, so they appear in no count.

Context Artifacts

Give the AI awareness of project knowledge beyond source code — database schemas, API specs, infrastructure c