The jCodeMunch MCP server provides token-efficient code exploration by retrieving source code using tree-sitter AST parsing. It targets code understanding tasks while reducing the amount of text sent to an LLM, with claims of 70+ language coverage and 86–99% token savings.
🛠️ Key Features
Token-efficient source code retrieval via tree-sitter AST parsing
Tree-sitter-based exploration to avoid reading entire files
70+ languages
Reported 86–99% token savings (96% average)
🚀 Use Cases
Code exploration for AI coding workflows (e.g., Claude, Cursor, Cline)
Managing limited context windows for source retrieval
Reducing LLM input costs during code research
⚡ Developer Benefits
Benchmark claims: 28.3× fewer tokens than a grep-and-read agent
“838B+ tokens saved” and “$4.2M+ in AI spend avoided” (as reported)
⚠️ Limitations
Data provided includes counter figures “as of 2026-08-17,” stated as floors, with unspecified current performance metrics beyond the excerpt.
The most token-efficient MCP server for precise source code retrieval via tree-sitter AST parsing. Cut AI token costs 86-99% on code exploration (96% average, benchmarked at 28.3x fewer tokens than a grep-and-read agent) and stop burning your context window reading entire files.
Real results, live from production838B+ tokens saved · 136,000+ reporting installs · $4.2M+ in AI spend avoided · 100,000+ kg CO₂ prevented
Counter figures as of 2026-08-17, valued at the $5/MTok Claude Opus input rate. All four only grow, so read them as floors. Live at jcodemunch.com.
Works with Claude Code, Cursor, VS Code, Codex CLI, Windsurf, Continue, and any MCP-compatible client.
Free for personal use. Use it to make money, and Uncle J. gets a taste. Fair enough? Commercial licenses below.
Our guarantee: if jCodeMunch doesn't pay for itself, you don't pay for jCodeMunch.
Why jCodeMunch?
Most AI agents explore repositories the expensive way: open entire files, skim thousands of irrelevant lines, repeat. That is not "a little inefficient." That is a token incinerator.
jCodeMunch indexes a codebase once and lets agents retrieve only the exact code they need: functions, classes, methods, constants, outlines, and tightly scoped context bundles, with byte-level precision. It parses source with tree-sitter, stores structured symbol metadata (signature, kind, qualified name, summary, byte offsets) alongside raw file content in a local index, and fetches exact implementations on demand instead of re-reading files over and over.
Measured with tiktoken cl100k_base across three public repos pinned to upstream commits, run 2026-09-03 on v1.108.316. Workflow: search_symbols (top 5) + get_symbol_source × 3 per query. Two baselines, same run, same corpus, same file reader:
Grep-top-3: rg -l the query terms, rank files by match count, open the top 3 whole. This is what a competent agent without the tool actually does, and it is the number to quote.
Read-all: every indexed source file concatenated. A ceiling nobody pays; retained for continuity with previously published figures.
Repository
Files
Symbols
Grep-top-3 baseline
jCodeMunch
vs grep
vs read-all
expressjs/express
186
455
15,724 avg
1,007 avg
15.6x
153.5x
fastapi/fastapi
1,186
13,240
85,296 avg
2,149 avg
39.7x
384.1x
gin-gonic/gin
98
1,451
31,975 avg
1,537 avg
20.8x
98.8x
Grand total (15 task-runs)
664,975
23,467
28.3x
241.1x
Against a grep-and-read agent: 96.5% reduction, 28.3x fewer tokens. No single multiple describes every query; the per-repo rows above are the spread. Against read-all the figure is 99.6%, but nobody pays that ceiling. Compact MUNCH wire encoding then trims a median 45.5% more bytes off responses.
50-iteration A/B test on a real Vue 3 + Firebase production codebase, jCodeMunch vs native tools (Grep/Glob/Read), Claude Sonnet 4.6, fresh session per iteration: success rate 80% vs 72%, timeout rate 32% vs 40%, mean cache creation down 10.5%. Tool-layer savings isolated from fixed overhead: 15-25%. One finding category appeared exclusively in the jCodeMunch variant: orphaned file detection via find_importers, a structural query native tools cannot answer without scripting. Full report: benchmarks/ab-test-naming-audit-2026-03-18.md
Mentioned by
Artur Skowroński (VirtusLab): "roughly 80% fewer tokens, or 5× more efficient — index once, query cheaply forever" · GitHub All-Stars #15
Traci Lim (AWS · ASEAN AI Lead): "structural queries that native tools can't answer: find_importers, get_blast_radius, get_class_hierarchy, find_dead_code" · 5 Repos That Save Token Usage in Claude Code
Julian Horsey (Geeky Gadgets): "3,850 tokens reduced to just 700 — a 5.5× improvement" · JCodeMunch AI Token Saver
No virtualenv to manage, nothing written into system Python, and it works as-is on PEP 668 distros (Ubuntu 24.04+, Debian 12+) where bare pip install is refused. Don't have uv yet?
init auto-detects your MCP clients (Claude Code, Claude Desktop, Cursor, Windsurf, Continue), writes their config entries, installs the CLAUDE.md prompt policy so your agent actually uses jCodeMunch, optionally installs enforcement hooks, optionally indexes your project, and audits your agent config files for token waste.
Other install paths
Command
Use it when
uvx jcodemunch-mcp
Zero install. Runs from an ephemeral environment — nothing lands on disk permanently. The client entries init writes already invoke the server this way, so for most setups this is all that ever runs. ⚠ Enforcement hooks are the exception: they're spawned by a minimal-PATH subshell and resolve the executable by name, so they need uv tool install (or pipx/pip) to work.
pipx install jcodemunch-mcp
You already standardise on pipx
pip install jcodemunch-mcp
Inside a virtualenv you manage yourself
Verify:
bash
jcodemunch-mcp --version
Manual Claude Code setup
bash
claude mcp add -s user jcodemunch -- uvx jcodemunch-mcp
No install step — uvx fetches and runs the server on demand. Prefer it on your PATH (and required for enforcement hooks)? uv tool install jcodemunch-mcp, then claude mcp add -s user jcodemunch jcodemunch-mcp.
Then tell the agent to prefer the tools. This matters more than people think; installation makes the tools available but does not break the agent's brute-reading habit. One line in your CLAUDE.md does it:
markdown
Call the jcodemunch_guide tool and strictly follow its instructions.
Using Cursor, Windsurf, Codex CLI, Antigravity, Gemini CLI, Qwen Code, Kiro, Cline, Zed, Goose, Hermes, Odysseus, or Paperclip? Every tested client configuration lives in CLIENTS.md. Optional extras (local semantic search, AI summaries per provider) are in QUICKSTART.md; the system surfaces each extra pulls in are documented in SECURITY.md.
Quickstart
Full walkthrough: QUICKSTART.md. The two-minute version, inside your agent after init:
Ask: "Index this repo with jcodemunch."
Ask: "Using jcodemunch, find the function that handles authentication and show me its source."
The agent should answer via search_symbols and get_symbol_source, returning tens of lines instead of whole files. Confirm with get_session_stats: it reports tokens served and savings for the session. That is where the numbers on the meter come from.
Want to skip initial indexing for popular frameworks? Pre-built starter packs: jcodemunch-mcp install-pack --list (free packs need no license).
What you can do
Retrieve one symbol instead of loading a file.get_symbol_source returns the exact function body, byte-precise, for the majority of edits that touch one function in a 700-line file (~95% savings on that read).
Assemble a whole task's context in one call.assemble_task_context classifies the task intent, extracts anchor symbols, and runs the right tool sequence under one token budget. plan_turn routes the turn before the first read.
Ask structural questions grep can't answer.find_importers, get_blast_radius, get_call_hierarchy, find_dead_code, get_changed_symbols, get_hotspots, search_ast anti-pattern sweeps, and more. Two of them sound alike and are not: check_references answers where a name is used (import sites plus every file whose content mentions it), find_references answers who imports it, over the import graph alone, so a call site is invisible to it.
Preflight risky changes, and know when to stop.check_edit_safe, check_delete_safe, get_pr_risk_profile, and plan_refactoring with edit-ready {old_text, new_text} blocks. The two safety checks return stop_rule.terminal: true means no further jcodemunch call moves the verdict, so re-running find_importers or check_references to be sure is wasted work. It means final, not safe. Hand the server your type checker's own output (jcodemunch-mcp import-trace --diagnostics <file>: mypy --output json, pyright --outputjson, tsc --pretty false, ruff --output-format json) and check_edit_safe, get_changed_symbols, get_pr_risk_profile and get_symbol_provenance say which symbols the checker already flags, as of which commit. Nothing runs a checker for you. False names the specific thing that would change the answer.
Trust the answers. Calibrated confidence scores, freshness flags, coverage contracts on absence claims, compiler-verified references via SCIP import, and automatic secret redaction before anything reaches the LLM.
Keep the index fresh automatically. Watch modes, agent hooks, and a VS Code extension close the staleness gap.
That's the highlight reel. The complete tour of 90+ tools, the MUNCH compact wire format, evidence receipts, offloadable-work annotation, and the session-economics instrumentation is in CAPABILITIES.md, with internals in UNDER_THE_HOOD.md.
What's new
v1.108.318 (2026-09-11) — the process is code that cannot skip a step, and the field is measured from result files
v1.108.317 (2026-09-04) — CI runs the harness on every change; publishing is a dispatched workflow
v1.108.316 (2026-09-02) — A display preference edited the data it was displaying
When does it help (and when doesn't it)?
Scenario
Native tool
jCodeMunch
Savings
Edit one function (700-line file)
Read → 700 lines
get_symbol_source → 30 lines
~95%
Understand a file's structure
Read → full content
get_file_outline → names + signatures
~80%
Find which file to edit
Grep many files
search_symbols → exact match
comparable
Edit requires whole-file context
Read → full content
get_file_content → full content
~0%
"What breaks if I change X?"
not possible
get_blast_radius
unique capability
It helps most on targeted edits (one function, one method, one class), which is the majority of real editing work. Edits that genuinely require the entire file (restructuring file-level state, reordering logic spanning hundreds of lines) see no advantage. Best fits: large repositories, unfamiliar codebases, agent-driven exploration, refactoring and impact analysis, and teams cutting AI token costs without making agents dumber.
Languages: 70+ via tree-sitter, including Python, JavaScript/TypeScript, Go, Rust, Java, C/C++, C#, PHP, Ruby, Swift, and Kotlin. Full matrix: LANGUAGE_SUPPORT.md. Monorepos: yes; incremental indexing, workspace-member detection, subpath scoping.
Deferring the tool schemas (Anthropic tool search)
If you reach jCodeMunch through the MCP connector on a model that supports tool search, you can keep our schemas out of your context prefix entirely and let Claude load only the two or three tools a request needs. You do not set defer_loading per tool — set it once for the whole server:
Send it with the beta header mcp-client-2025-11-20. Both halves are required — mcp_servers alone is a validation error, and so is mcp_toolset without the matching mcp_server_name.
⚠ The MCP connector takes a URL, so this applies to jCodeMunch served over sse or streamable-http (jcodemunch-mcp serve --transport streamable-http), not to the default local stdio setup. On stdio, whether schemas are deferred is up to your client, and tool_surface: "counter" below is the lever you control.
The configs block above follows Anthropic's own advice — keep your 3–5 most-used tools resident so common requests skip the search round trip — and per-tool configs overrides default_config.
Deferred definitions are excluded from the system-prompt prefix and appended inline as tool_reference blocks when Claude discovers them, so prompt caching is preserved — this is not the cache-invalidating kind of dynamic tool list. At least one tool in the request must stay non-deferred, or the API returns a 400.
⚠ This is a different mechanism from our own tool_surface: "counter", and you do not need both. Tool search is host-side and works across every MCP server you have connected; the Counter is server-side, works on any host including ones with no tool-search support, and is what init configures on a first-ever install. Pick whichever your host supports — see CONFIGURATION.md for the Counter and jcodemunch-mcp surface for what your install actually advertises.
Security, privacy, and background behavior
Local-first by design: indexes live at ~/.code-index/, and the base package's only default network behavior is an anonymous savings counter (random ID plus aggregate token counts, no code, no paths, no PII; opt out with share_savings: false). Everything the server does beyond answering a tool call (file watching, the opt-in login service, license validation, model downloads, org reporting) is opt-in or opt-out, visible, and reversible, and every item is enumerated in SECURITY.md alongside the path-traversal, symlink, and secret-redaction controls.
Grammar pack override (#608). The dependency tree-sitter-language-pack is pinned <1.0.0 because the 0.x wheels bundle every grammar and parsing stays local. You can override it with pip install -U tree-sitter-language-pack after installing; nothing in the code refuses it. What you accept, measured against 1.17.0 on 2026-09-11: the 1.x pack ships no grammars and fetches each one over the network into its cache directory (measured on Windows: %LOCALAPPDATA%\tree-sitter-language-pack\v<version>\libs; jcodemunch-mcp install-status prints the path on any platform) the first time a language is parsed, so an airgapped install parses nothing, and the nim grammar changed upstream, so nim files yield no symbols (the manifest also lacks autohotkey, ejs and verse, which costs nothing here: those three are parsed by jCodeMunch's own extractors, not by tree-sitter). An install on a 1.x pack says so: every index_folder result carries a grammar_pack block and a warning naming the version, the cache directory and each language whose grammar failed, and jcodemunch-mcp install-status prints the same. Dropping the pin is a separate decision that needs the offline story first.
Per-project configuration
Most settings live in the global ~/.code-index/config.jsonc, but any of them can be overridden for a single repository by dropping a .jcodemunch.jsonc at its root. It is an overlay: keys it declares win, keys it omits fall through to global and then to the built-in default, so it only needs to contain what differs.
Racket projects routinely define their own defining forms with define-syntax, and a static parser cannot know what those bind — (defstep (check-admin) ...) is indistinguishable from a function call. Declaring them makes their bindings searchable:
Each entry maps a form name to what it binds: function, constant, class or type. Where the name sits is read from the source rather than declared — (defstep (check-admin) ...) takes the head of the parameter list, (defstudy consent ...) takes the bare symbol — so a form that appears in both shapes works either way.
⚠ This is an assertion, not something jCodeMunch can verify. A wrong declaration puts a name in the index that Racket does not actually bind. Declarations are also matched only after every built-in form, so declaring define or struct has no effect — the built-in handling wins.
Declaring what a Racket #lang looks like
A #lang line names a reader, and jCodeMunch's Racket parser reads S-expressions. The distribution's langs are built in (racket/*, typed/racket*, s-exp, info, at-exp …, and the document langs scribble/*, pollen, punct, markdown …), but a project's own lang is unknown to it and is treated as a document — no symbols, still text-searchable — until you say what its syntax is:
sexp is plain S-expressions; at-exp is at-exp text bodies over Racket (read with @ as the command character, exactly as #lang at-exp reads them, so prose containing ;"# or | is prose); text is a document language that is never walked. A key also covers its sub-langs (conscript matches conscript/with-require), and a project may demote a lang as well as promote one. An at-exp lang whose reader uses another command character declares it with the object form — "mylang": {"tier": "at-exp", "command_char": "◊"} — the way Racket's make-at-readtable takes #:command-char.
Both keys change what the parser emits for unchanged files, so a change to either is stamped on the index and forces one full re-parse on the next index (rebuild_reason: "racket_config_changed"); you do not need to touch the files or clear the index. An index holding Racket files that was built before this stamp existed re-parses once the same way (rebuild_reason: "racket_index_predates_gate").
jCodeMunch-MCP is released under the jCodeMunch-MCP Dual-Use License (full terms). Free for non-commercial use. Commercial use requires a paid license, one-time, sold by jMunch LLC via Stripe:
Not sure it's worth it? Run your own numbers through the ROI calculator, or forward the finance-team version to whoever signs off. The guarantee stands: if jCodeMunch doesn't pay for itself, you don't pay for jCodeMunch.
Conditions on all uses: retain the copyright notice, clearly mark modifications and keep the original author's name intact (he's kinda full of himself), and include a prominent modification notice in source redistributions. The Software may not be renamed, rebranded, or published to any public package registry, and is provided "AS IS" without warranty. LICENSE controls.
FAQ
How much can I save on Claude / Opus tokens?
In retrieval-heavy workflows, code-reading tokens typically drop 86-99%, benchmarked at 96.5% average (28.3x) against a grep-and-read agent across 15 tasks and 3 repositories. Per-query results span 7.6x to 81.2x. Methodology: TOKEN_SAVINGS.md and benchmarks/.
How is this different from RAG or grep-based tools?
jCodeMunch retrieves at the symbol level with byte-level precision (functions, classes, importers, blast radius, hierarchies) rather than fuzzy chunks (RAG) or raw line matches (grep) the agent still has to read and reason over.
Is it free for personal use?
Yes. Commercial use needs a license; see above.
