Serveur MCP CodeGraphX

CodeGraphX (CGX)

CodeGraphX (CGX) is a local, token-efficient codebase graph engine for AI coding agents and human developers. It parses your source with Tree-sitter, stores a bi-temporal semantic graph in SQLite, and serves it over a CLI and an MCP server — so an agent can ask "what breaks if I change this function?" and get an exact answer in a few hundred tokens instead of reading 50 files.

---

What is CodeGraphX?

Definition — CodeGraphX is a code intelligence layer: a persistent, queryable graph of your codebase's symbols (functions, classes, variables), their relationships (calls, called_by, imports, inherits, cross-language API_CALLS), and how they change over time.

What kind of tool is it? It sits in the same family as a Language Server (LSP) or a code-search index, but it is purpose-built for LLM agents:

• A Language Server answers "go to definition" for an editor, one symbol at a time.
• A grep / embedding search finds text, not structure.
• CodeGraphX answers structural, whole-repo reasoning questions — impact/blast-radius, dependency cycles, cross-language contracts, "does this symbol even exist?" — and serializes the answer in a token-optimized form an agent can drop straight into its context.

The core problem it solves — AI coding agents burn most of their token budget re-discovering a codebase: opening files, scrolling, grepping, re-reading the same modules every session. CodeGraphX indexes once and answers those questions from a graph, so the agent spends tokens reasoning instead of scanning.

---

Why it's useful (the token economics)

Every agent action has a token cost. Consider a common question: "What will break if I change validateInput?"

Approach	What the agent does	Rough context cost
No CGX	Greps for validateInput, opens ~15–50 candidate files, reads each to trace call sites	tens of thousands of tokens, several tool round-trips
With CGX	explain_impact({ symbol_name: "validateInput" }) → one small JSON of upstream callers + downstream callees	a few hundred tokens, one round-trip

The savings compound because CGX serializes its graph in TOON (Token-Oriented Object Notation) instead of JSON — a compact tabular encoding that removes the repeated keys and braces that dominate JSON token counts.

Honest framing: the accuracy numbers below are measured against a hand-labeled corpus. The token figures above are an illustrative model of typical agent behavior, not a lab benchmark — the exact ratio depends on your repo and agent. The mechanism (one targeted graph query vs. many file reads) is what delivers the win.

Where CGX pays off most:

• Large / unfamiliar repos where "read everything" is infeasible.
• Long agent sessions (the graph is a cache the agent doesn't have to rebuild).
• Cross-language stacks (React ⇄ Express/Flask/FastAPI) where call graphs don't span files.
• Refactors and reviews, where blast-radius and dead/broken-import detection matter.

---

Full capabilities

Capability	What it gives you
🧠 Symbol graph	Every function/class/method with calls, called_by, imports, inherits edges
⚡ O(1) symbol lookup	Bloom filter answers probable_yes / definite_no without touching the DB
💥 Impact tracing	Recursive upstream (callers) / downstream (callees) blast-radius via a SQL CTE
🌉 Cross-language linking	Matches frontend fetch/axios calls to backend routes as confidence-scored API_CALLS edges
🕑 Bi-temporal history	Append-only graph — query the codebase as of any commit; nothing is ever destroyed
🩺 Doctor diagnostics	Reports missing/unresolvable imports, unresolved call targets, circular dependencies, syntax errors
✅ Task verification	Compares a task description against a commit's actual symbol changes; flags untested additions
🔀 Session / branch diff	Structural summary of added/removed/modified symbols vs. HEAD or a branch
🌐 Interactive dashboard	D3.js force graph of the whole codebase in the browser
📦 TOON artifacts	Token-optimized graph + file-index exports for agent context injection
🔐 100% local	No cloud, no telemetry, no network — code never leaves the machine
🤖 MCP server	6 tools exposed to any MCP-compatible agent, with zero-config auto-indexing

---

Tech used in this project

Layer	Technology
Parsing	Tree-sitter with per-language grammars (tree-sitter-python, -javascript, -typescript, -html, -css)
Storage	SQLite via better-sqlite3 (default) or Postgres (pg), through the Kysely type-safe query builder
Agent protocol	Model Context Protocol SDK (@modelcontextprotocol/sdk) over stdio
Probabilistic lookup	bloom-filters
Token-optimized output	TOON (@toon-format/toon)
CLI	Commander + @clack/prompts for the interactive setup
File watching	chokidar
Runtime	Node.js ≥ 18, CommonJS

---

Supported languages

Language	Extensions	Extracts
Python	.py	functions, classes, calls, imports (incl. package-relative . / ..), Flask/FastAPI routes
JavaScript	.js, .jsx	functions, classes, arrow fns, calls, imports/require, local bindings, fetch/axios calls, Express routes
TypeScript	.ts, .tsx	same as JS plus TS-specific declarations
HTML	.html	elements / structural symbols
CSS	.css	selectors (classes, ids)

---

Installation

# Global (recommended)
npm install -g codegraphx

# Or per-project
npm install --save-dev codegraphx

# Verify
codegraphx --version        # 1.2.4

The CLI is available as codegraphx, cgx, and the MCP entrypoint cgx-mcp.

---

Quick start

cd your-project

# 1. Index the codebase → writes .codegraphx/ + .codegraphx.db
cgx init

# 2. Ask questions
cgx query authenticateUser                       # where is it, what calls it, what it calls
cgx impact authenticateUser --direction downstream --depth 5
cgx impact authenticateUser --direction upstream
cgx stats                                        # files / symbols / edges
cgx doctor                                       # health report

# 3. Optional live view
cgx watch                                        # re-index on file change
cgx dashboard                                    # open the D3 graph in a browser

Re-scanning is a full rewrite. Each cgx init / cgx scan rebuilds .codegraphx/ from scratch: files deleted from disk are evicted from the graph, stale artifacts are wiped, and the database temporal-closes removed symbols — so the graph always matches your working tree. Unchanged files are still served from cache, so it stays fast.

CLI command reference

Command	Purpose
cgx setup	Wire the MCP server + skill into your coding CLIs (see below)
cgx init / cgx scan	Full index of the codebase
cgx query <symbol>	Show a symbol's file, type, calls, and called_by
cgx impact <symbol> [--direction up/downstream] [--depth N]	Trace the blast radius
cgx doctor [--json] [--strict] [--no-calls]	Diagnostics: imports, calls, cycles, syntax
cgx stats	Counts of files / symbols / functions / classes / edges
cgx watch	Auto-update the graph on file changes
cgx dashboard	Open the interactive codegraph.html
cgx git-hook <install|remove>	Auto-scan on post-commit / pre-push
cgx session summary [--branch <b>]	Structural change summary for the current session
cgx verify --task <desc> --commit <hash>	Task-vs-commit verification evidence

---

MCP server (for AI agents)

CodeGraphX ships an MCP stdio server exposing 6 tools. Zero-setup: you don't have to scan first — on its first start in a project, the server indexes the codebase in the background. While indexing, get_graph_status reports "indexing"; once "ready", all tools are live.

Available MCP tools

Tool	Description	Parameters
get_graph_status	Readiness: indexing / ready / error, plus file count	—
list_files	List indexed files	filter?: string
check_symbol_exists	O(1) Bloom lookup → probable_yes / definite_no	name: string
explain_impact	Upstream callers + downstream callees of a symbol	symbol_name: string
verify_task	Compare a task description to a commit's real changes	task_description: string, commit_hash?: string
get_session_diff	Structural summary of the current session/branch	branch?: string (default HEAD)

Example agent workflow

User: "What breaks if I change the validateInput function?"

Agent (via MCP):
  1. check_symbol_exists({ name: "validateInput" })   → { exists: "probable_yes" }
  2. explain_impact({ symbol_name: "validateInput" })
       → { used_by_upstream: ["src/auth.js::login"],
           breaks_downstream: ["src/api.js::handleRequest"] }

Result: exact answer, no file scanning.

Picking the project root

The server indexes the directory it starts in. If your client doesn't set one, pass it explicitly:

cgx-mcp --project-root /path/to/project
# or
CGX_PROJECT_ROOT=/path/to/project cgx-mcp

---

Setup — how easy is it?

One command

cgx setup

It auto-detects the coding CLIs you have installed, lets you multi-select which to configure, and wires each one — no hand-editing config, no hunting for absolute paths. It registers the MCP server (via each CLI's native command where available, with a JSON-file fallback) using an absolute Node path + the bundled cgx-mcp, so it works for both global and local installs, and it installs the CGX usage skill/instructions in each CLI's format.

cgx setup                                   # interactive multi-select
cgx setup --agents claude,gemini --yes      # non-interactive (CI / scripted)
cgx setup --project                         # register for the current repo only

Registration is user/global by default, so you run it once and it works in every project — cgx-mcp resolves the project from wherever your CLI launches. Existing config is preserved; a backup is written before the first edit.

Supported coding-agent CLIs

Agent	Detected via	MCP registration	Skill / instructions
Claude Code	claude on PATH or ~/.claude	claude mcp add → ~/.claude.json or .mcp.json	~/.claude/skills/cgx/SKILL.md
Gemini CLI	gemini on PATH	~/.gemini/settings.json or .gemini/settings.json	GEMINI.md (context file)
Antigravity CLI	agy on PATH	~/.gemini/config/mcp_config.json or .agents/mcp_config.json	~/.gemini/skills/cgx/SKILL.md or .agents/skills/cgx/SKILL.md
OpenCode	opencode on PATH	opencode mcp → config file	AGENTS.md
Cursor	~/.cursor present	~/.cursor/mcp.json or .cursor/mcp.json	.cursor/rules/*.mdc

Then open your CLI in any project and say "use cgx to explore this codebase."

If cgx setup doesn't work — manual setup

The MCP entry is always the same shape: run the bundled cgx-mcp with an absolute path to node (some clients don't inherit your shell PATH). Find them with:

which node                    # e.g. /usr/local/bin/node
which cgx-mcp                 # e.g. /usr/local/lib/node_modules/codegraphx/bin/cgx-mcp

Claude Code — from your project:

claude mcp add codegraphx -- npx -y -p codegraphx cgx-mcp

or .mcp.json:

{ "mcpServers": { "codegraphx": { "command": "npx", "args": ["-y", "-p", "codegraphx", "cgx-mcp"] } } }

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS). No project dir, so set the root:

{ "mcpServers": { "codegraphx": {
  "command": "npx",
  "args": ["-y", "-p", "codegraphx", "cgx-mcp", "--project-root", "/path/to/project"]
} } }

Gemini CLI — .gemini/settings.json (use an absolute node path):

{ "mcpServers": { "codegraphx": {
  "command": "/ABSOLUTE/PATH/TO/node",
  "args": ["/ABSOLUTE/PATH/TO/node_modules/codegraphx/bin/cgx-mcp"],
  "cwd": "/ABSOLUTE/PATH/TO/PROJECT"
} } }

Antigravity CLI (agy) — file-only, no native command. Global ~/.gemini/config/mcp_config.json or per-project .agents/mcp_config.json:

{ "mcpServers": { "codegraphx": {
  "command": "/ABSOLUTE/PATH/TO/node",
  "args": ["/ABSOLUTE/PATH/TO/node_modules/codegraphx/bin/cgx-mcp"]
} } }

Cursor / Windsurf / other MCP clients — ~/.cursor/mcp.json (or project .cursor/mcp.json), same structure; point cwd at the project root.

Verify the connection:

# inside the CLI
/mcp list           # should show: codegraphx — Connected (6 tools)

# or drive the server directly
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' | npx -y -p codegraphx cgx-mcp

Still disconnected?

1. Use an absolute path to node, not npx.
2. Make sure cwd / --project-root matches your project exactly.
3. gemini trust if using project-scoped Gemini settings.
4. Inspect stderr: node /path/to/cgx-mcp 2>&1 | head -20.

---

Architecture

Data flow

Source files
   → parser.js            Tree-sitter, one adapter per language
   → entities.js          SymbolEntity / FileEntity / EdgeEntity (the domain model)
   → edgebuilder.js       CALLS / IMPORTS / INHERITS edges
   → cross-language-linker.js   fetch/axios ⇄ Express/Flask/FastAPI  →  API_CALLS edges
   → scanner.js           orchestrates a full scan; writes DB + .codegraphx/ artifacts
   → store/sql-store.js   Kysely ORM — append-only bi-temporal graph (SQLite / Postgres)

After a scan, .codegraphx.db (in the project root) is the source of truth. The .codegraphx/ directory holds derived artifacts: codegraph.html (dashboard), codegraph-graph.json (D3), codegraph.toon + file_index.toon (token-optimized), symbols.bloom, and cache.json.

Bi-temporal, append-only storage

Rows in files, symbols, and edges are never deleted. Each carries valid_from_commit_id and valid_to_commit_id; NULL in the latter means "currently active." Editing a symbol closes the old row and inserts a new one; deleting a file temporal-closes its rows. This lets CGX answer "what did the graph look like at commit X?" and powers session/branch diffs and task verification — without ever losing history.

Key modules

Module	Role
src/parser.js	Routes a file to its language adapter; returns declaredSymbols, imports, calls
src/languages/*	Per-language Tree-sitter adapters (python, javascript, typescript, html, css)
src/entities.js	SymbolEntity, FileEntity, EdgeEntity, Snapshot domain model
src/edgebuilder.js	Builds structural edges from parsed references
src/cross-language-linker.js	Confidence-scored API_CALLS edges across the frontend/backend boundary
src/resolver.js	Canonical import resolver (JS .//../, Python package-relative, dotted paths)
src/scanner.js	runScan() — the indexing entrypoint; full-rewrite semantics; mcpMode skips heavy outputs
src/store/sql-store.js	SqlGraphStore — all DB reads/writes; traceImpact uses a recursive SQL CTE; pruneDeletedFiles
src/store.js	GraphStore — in-memory + cache.json layer used by the scanner and doctor
src/doctor.js	Diagnostics engine (imports / calls / cycles / syntax)
src/verifier.js	Task-vs-commit verification evidence
src/sdk/index.js	IntelligenceSDK — programmatic API for embedding CGX
src/server/mcp-server.js	CodeGraphXServer — the MCP stdio server
src/setup/*	cgx setup orchestrator + per-CLI adapters
src/db/*	Kysely db instance + migrations (DB_DIALECT=postgres + DATABASE_URL to switch)

---

Benchmarks & accuracy

CGX is meant to be trusted in place of reading code, so its graph is scored against a hand-labeled golden corpus (tests/golden/) where every symbol, edge, cross-language link, and import cycle is known. The same harness gates CI (tests/golden/accuracy.test.js) — a parser regression fails the build.

Extraction accuracy (curated corpus: 3 fixtures, 9 files, 20 symbols):

Category	Precision	Recall	F1
Symbols	100%	100%	100%
Structural edges (CALLS / IMPORTS / INHERITS)	100%	100%	100%
Cross-language API links	100%	100%	100%
Endpoint tagging	100%	100%	100%

Reasoning & determinism:

Check	Result
Impact tracing (exact reachable set)	4/4 (100%)
Circular-import detection (recall)	1/1 (100%)
Circular-import false positives	0
Deterministic across re-scans	yes

Throughput: ~82 files/sec, ~183 symbols/sec on the corpus (full scan in ~110 ms).

Reproduce and regenerate BENCHMARK.md + benchmark-results.json:

npm run benchmark

These numbers verify extraction correctness on a controlled corpus, not coverage of every language construct. Extend the bar by adding a fixture under tests/golden/<name>/ with a ground-truth.json.

---

Cross-language intelligence

During a scan, CGX extracts the HTTP requests your client makes and the routes your server exposes, then matches them into confidence-scored API_CALLS edges:

fetch('/api/users')        ──API_CALLS(0.9)──▶  app.get('/api/users', listUsers)   [Express]
axios.post('/api/orders')  ──API_CALLS(0.9)──▶  @router.post('/api/orders')        [FastAPI]
fetch(`/api/users/${id}`)  ──API_CALLS(0.7)──▶  @app.route('/api/users/<id>')      [Flask]

• Frontend: fetch(...) (with method), axios.get/post/..., axios({ url, method }).
• Backend: Express/router, Flask (@app.route(..., methods=[...])), FastAPI (@router.get, @app.post, …).
• Confidence: 0.9 exact path+method, 0.75 exact path/other method, 0.7 parameterized path+method, 0.55 parameterized/other. Path params (:id, {id}, <int:id>) are normalized first, so /api/users/${id} links to /api/users/{user_id}.

Route handlers are tagged with an endpoint ontology marker, and explain_impact traverses API_CALLS edges — so "what calls this backend handler?" returns the frontend callers across the language boundary.

---

Security & privacy

• ✅ 100% local — no network, no telemetry, no cloud sync.
• ✅ Read-only — CGX never modifies your source.
• ✅ Static only — code is parsed, never executed.
• ✅ Configurable ignores — exclude sensitive paths via .codegraphxrc.

The MCP server runs with your terminal's permissions and only reads files matching the configured extensions/ignores. For sensitive projects, run it in a sandbox or container.

---

Configuration

Project config lives in .codegraphxrc (JSON) or codegraphx.config.json:

{
  "extensions": [".py", ".js", ".ts", ".jsx", ".tsx", ".html", ".css"],
  "ignore": [".git", "node_modules", "__pycache__", ".venv", "dist", "build", "coverage"],
  "outputDir": ".codegraphx",
  "outputFile": "codebase.json",
  "bloomErrorRate": 0.01
}

Add .codegraphx/ and .codegraphx.db to your .gitignore — they're build artifacts.

---

FAQ

Do I need to run a scan every time? No — the MCP server auto-indexes on first use. Re-scan (or cgx watch) to refresh after changes. Every re-scan is a clean full rewrite, so deletions and renames never leave stale graph data.

Large codebases? Yes — incremental parsing means unchanged files are served from cache; only changed files are re-parsed.

Private / proprietary code? Fine — everything runs locally, nothing leaves your machine.

MCP shows "Disconnected"? Use an absolute node path, confirm cwd/--project-root, gemini trust for project scope, and check stderr (see manual setup above).

How accurate is the Bloom filter? Tunable via bloomErrorRate (default 0.01). False positives only trigger a fallback lookup — never false negatives.

---

Contributing

git clone https://github.com/techcraze00/CodeGraphX.git
cd codegraphx
npm install
npm link          # exposes `codegraphx` / `cgx` globally
npm test

Adding a language: add the Tree-sitter grammar to package.json, create an adapter under src/languages/<lang>/, register it in src/languages/index.js, and add a golden fixture under tests/golden/.

---

License

MIT © 2026 Prayas Jadhav. See LICENSE.

---

Links

• Website / docs: techcraze00.github.io/CodeGraphX
• Package: codegraphx on npm
• Repository: github.com/techcraze00/CodeGraphX
• Issues: github.com/techcraze00/CodeGraphX/issues

Acknowledgments

Tree-sitter · Model Context Protocol · Kysely · TOON · bloom-filters