code-graph-mcp

A high-performance code knowledge graph server implementing the Model Context Protocol (MCP). Indexes codebases into a structured AST knowledge graph with semantic search, call graph traversal, and HTTP route tracing — designed to give AI coding assistants deep, structured understanding of your code.

Features

• Multi-language parsing — Tree-sitter AST extraction across tiers of depth:

  • Full (calls + imports + inheritance): TypeScript/TSX, JavaScript, Go, Python, Rust, Java. HTTP route extraction additionally covers TypeScript/TSX + JavaScript (Express/Connect), Go (net/http), Python (Flask/FastAPI), and Rust axum (.route() chains with inline .nest() prefixes; named handlers only) — actix/rocket and Java Spring are not yet route-extracted

  • Smoke-tested (calls + imports + inheritance): C#, Kotlin, Ruby, PHP, Swift, Dart

  • Limited (functions + calls + #include imports + gtest test markers + C++ base-class inheritance + Class::method scope qualification): C, C++

  • Scripting: Bash (functions + commands + source/. imports), Markdown (headings)

  • File-FTS only (no AST symbol extraction): HTML, CSS, JSON

  • Test markers — detected from the AST for Rust (#[test] / #[cfg(test)]), JavaScript/TypeScript/TSX (describe/it/test blocks and their callbacks) and C/C++ (gtest TEST macros). Go, Python, Java and the rest fall back to path/name heuristics (tests/, src/test/java/, *_test.*, *.test.js, test_*, *Test)

  • Value/type references — a references relation for symbols that are used without being called, imported or inherited (path-qualified constants, type-position uses, functions passed as values): Rust, TypeScript/TSX, JavaScript, Python, Go, Java, C, C++

• Semantic code search — Hybrid BM25 full-text + vector semantic search with Reciprocal Rank Fusion (RRF), powered by sqlite-vec

• Call graph traversal — Recursive CTE queries to trace callers/callees with cycle detection

• HTTP route tracing — Map route paths to backend handler functions (Express, Flask/FastAPI, Go net/http, Rust axum)

• Dead code detection — Find unreferenced symbols with smart Orphan/Exported-Unused classification

• Impact analysis — Determine the blast radius of code changes by tracing all dependents

• Incremental indexing — Merkle tree change detection with file system watcher for real-time updates. Smart event filtering skips metadata-only changes (chmod, xattr)

• Context compression — Token-aware snippet extraction for LLM context windows (L0→full code, L1→summaries, L2→file groups, L3→directory overview). Compact JSON output saves 15-20% tokens

• Embedding model — Optional local embedding via Candle (feature-gated embed-model). Context reordered to prioritize structural relations over code for better embedding quality

• Self-healing — Automatic SQLite corruption recovery with rebuild. Startup repair for incomplete indexing (Phase 3 failures)

• MCP protocol — JSON-RPC 2.0 over stdio, plug-and-play with Claude Code, Cursor, Windsurf, and other MCP clients

• Claude Code Plugin — First-class plugin with skills (explore, index), a code-explorer agent, auto-indexing hooks, StatusLine integration, and self-updating

Related MCP server: codebase-memory-mcp

Why code-graph-mcp?

Unlike naive full-text search or simple AST dumps, code-graph-mcp builds a structured knowledge graph that understands the relationships between symbols across your entire codebase.

Incremental by Design

BLAKE3 Merkle tree tracks every file's content hash. On re-index, only changed files are re-parsed — unchanged directory subtrees are skipped entirely via mtime cache. When a function signature changes, dirty propagation automatically regenerates context for all downstream callers across files.

Hybrid Search, Not Just Grep

Combines BM25 full-text ranking (FTS5) with vector semantic similarity (sqlite-vec) via Reciprocal Rank Fusion (RRF) with raw score blending — so searching "handle user login" finds the right function even if it's named authenticate_session. Results are auto-compressed to fit LLM context windows.

Scope-Aware Relation Extraction

The parser doesn't just find function calls — it tracks them within their proper scope context. Extracts calls, imports, inheritance, interface implementations, exports (ESM export and CommonJS module.exports = { … } / exports.name = …), value/type references, and HTTP route bindings. Same-file targets are preferred over cross-file matches to minimize false-positive edges.

HTTP Request Flow Tracing

Unique to code-graph-mcp: trace from GET /api/users → route handler → service layer → database call in a single query. Supports Express/Connect, Flask/FastAPI, Go net/http, and Rust axum.

Zero External Dependencies at Runtime

Single binary, embedded SQLite, bundled sqlite-vec extension, optional local embedding model via Candle — no database server, no cloud API, no Docker required. Runs entirely on your machine.

Built for AI Assistants

Every design decision — from token-aware compression to node_id-based snippet expansion — is optimized for LLM context windows. Works out of the box with Claude Code, Cursor, Windsurf, and any MCP-compatible client.

Performance

Every row below is a line code-graph-mcp benchmark prints, measured on this repository (278 files, 5,065 nodes, 10,731 edges) with a release build. Run the same command on your own project — the numbers that matter are yours, and these scale with tree size and machine.

Efficiency: code-graph vs Traditional Tools

How many TOOL CALLS each question takes, comparing code-graph-mcp against Grep

• Read + Glob. These are call counts, not token measurements: the left column is the round trips the traditional approach needs before it can answer, the right column is the single call that answers it. Token cost follows call count only loosely — how much source lands in the context window depends on the files, so this table does not claim a token ratio.

Tool Call Reduction

Session token savings

tests/effectiveness_bench.rs is the only reproducible number here, and it is worth being precise about what it measures. For each of five navigation tasks it runs the real CLI on a fixture project and compares the response size in bytes against a hand-set baseline_bytes for the Grep+Read approach. Bytes stand in for tokens; the baselines are estimates committed once and held fixed, so its value is regression tracking over releases, not a measurement of your project.

cargo build --no-default-features
cargo test --test effectiveness_bench --no-default-features -- --ignored --nocapture

On the fixture at v0.123.0 that prints 942 / 23000 = 0.04x, and the test fails if the overall ratio ever exceeds 0.60. Earlier revisions of this README carried an "Overall Session Efficiency" table (~80% fewer calls, ~95% fewer lines read, 40-60% session savings) whose first rows had no source at all; they are gone rather than dressed up.

What code-graph Uniquely Enables

• Impact analysis — "Changing conn affects 33 functions across 4 files, 78 tests at HIGH risk" — impossible to derive manually with Grep

• Transitive call tracing — Follow main → run_serve → handle_message → handle_tools_call → conn in one query

• Semantic search — Find authenticate_session when searching "handle user login"

• Dependency strength — Not just "file A imports file B", but "file A uses 38 symbols from file B"

When Traditional Tools Are Still Better

Architecture

src/
├── domain.rs     # Shared constants, relation types, env-var config
├── resolve.rs    # Shared symbol resolution + ambiguity verdicts (CLI and MCP)
├── outcome.rs    # Retrieval-adoption metrics from session transcripts
├── cli/          # Every `code-graph-mcp <cmd>` subcommand (one file per command)
├── mcp/          # MCP protocol layer (JSON-RPC, tool registry, server)
│   └── server/   # McpServer with IndexingState + CacheState sub-structs
├── parser/       # Tree-sitter parsing, relation extraction, LanguageConfig dispatch
├── indexer/      # 3-phase pipeline, Merkle tree, file watcher
├── storage/      # SQLite schema (v10), CRUD, FTS5, migrations
├── graph/        # Recursive CTE call graph queries
├── search/       # RRF fusion search combining BM25 + vector
├── embedding/    # Candle embedding model (optional, masked mean pooling)
├── snapshot/     # Portable graph snapshots (create / verify / install)
├── sandbox/      # Context compressor with token estimation
└── utils/        # Language detection, config

Installation

Option 1: Claude Code Plugin (Recommended)

Install as a Claude Code plugin for the best experience — includes skills, the code-explorer agent, auto-indexing hooks, StatusLine health display, and automatic updates:

# Step 1: Add the marketplace
/plugin marketplace add sdsrss/code-graph-mcp

# Step 2: Install the plugin
/plugin install code-graph-mcp

What you get:

• MCP Server — All code-graph tools available to Claude

• Skills — explore (structure-first navigation before reading files) and index (health-check / re-index / full rebuild); see Plugin Skills

• Code Explorer Agent — Deep code understanding expert via code-explorer

• Auto-indexing Hook — Incremental index on every file edit (PostToolUse)

• StatusLine — Real-time health display (nodes, files, watch status) — compatible with other plugins' StatusLine via composite multiplexer

• Auto-update — Checks for a new version at session start (throttled to at most one check every 2 minutes). Between forced checks the re-check interval is 30 minutes after an "up to date" answer and 6 hours while an update is already pending. Updates install silently.

Manual Update

npm update -g @sdsrs/code-graph

Then reconnect the MCP server in Claude Code with /mcp.

Note: Auto-update is disabled in the source repo directory (dev mode). Use manual update when developing the plugin itself.

Turning auto-update off

Set it in the env block of ~/.claude/settings.json (the environment of the process hosting the MCP server), then reconnect with /mcp.

An update that keeps failing also stops hammering: after 5 failed attempts at the same release the updater drops to one retry per day (and retries immediately when a newer release is published), instead of re-downloading on every session. While it is in that state the statusline shows ⚠ update stuck and code-graph-mcp doctor prints the manual update command.

Invited-memory mode (quieter prompts)

By default, every user prompt the plugin deems code-related gets a small context injection from code-graph CLI output. If you'd rather rely on MEMORY.md + explicit tool calls, opt into invited-memory mode:

1. Adopt the plugin contract into your project (idempotent, self-heals):

   code-graph-mcp adopt

   This writes a sentinel-wrapped managed block into <cwd>/CLAUDE.md (auto-loaded each session) plus <cwd>/.claude/plugin_code_graph_mcp.md (the full decision table, opened on demand). Run code-graph-mcp unadopt to remove both; content outside the managed block is kept.

2. Set the activation env var in ~/.claude/settings.json:

   {
     "env": { "CODE_GRAPH_QUIET_HOOKS": "1" }
   }

3. Restart Claude Code. Session startup skips the project-map injection, UserPromptSubmit stops auto-injecting context, and the MCP instructions become a short pointer to the MEMORY.md file.

Option 2: Claude Code MCP Server Only

Register as an MCP server without the plugin features:

claude mcp add code-graph-mcp -- npx -y @sdsrs/code-graph

Option 3: Cursor / Windsurf / Other MCP Clients

Add to your MCP settings file (e.g. ~/.cursor/mcp.json):

{
  "mcpServers": {
    "code-graph": {
      "command": "npx",
      "args": ["-y", "@sdsrs/code-graph"]
    }
  }
}

Option 4: npx (No Install)

Run directly without installing:

npx -y @sdsrs/code-graph

Option 5: npm (Global Install)

Install globally, then run anywhere:

npm install -g @sdsrs/code-graph
code-graph-mcp

Uninstallation

Claude Code Plugin

# Uninstall the plugin
/plugin uninstall code-graph-mcp

# (Optional) Remove the marketplace
/plugin marketplace remove code-graph-mcp

# (Optional) Clean up all config and cache data
node ~/.claude/plugins/cache/code-graph-mcp/code-graph-mcp/*/scripts/lifecycle.js uninstall

Claude Code MCP Server

claude mcp remove code-graph-mcp

Cursor / Windsurf / Other MCP Clients

Remove the code-graph entry from your MCP settings file (e.g. ~/.cursor/mcp.json).

npm (Global)

Run the teardown first, while the CLI still exists — npm 7 removed the preuninstall/postuninstall lifecycle scripts, so nothing runs on your behalf during npm uninstall. Doing it in the other order leaves the hook entries in ~/.claude/settings.json and the ~40 MB binary cache behind, with no command left on disk to remove them.

code-graph-mcp uninstall     # restore statusline, strip hooks, drop the cache
npm uninstall -g @sdsrs/code-graph

MCP Tools

tools/list advertises exactly these seven. Several older niche tools were folded into flags on them, so one call now covers what used to take a separate tool:

Hidden aliases. These names are not in tools/list but still dispatch via tools/call, so existing clients keep working: trace_http_chain / find_http_route (→ get_call_graph with route_path), read_snippet (→ get_ast_node), dependency_graph, find_similar_code, find_dead_code, plus the management tools start_watch, stop_watch, get_index_status and rebuild_index. impact_analysis is removed — calling it returns Unknown tool; use get_ast_node with include_impact=true, or the CLI's impact --json for the full report.

CLI Commands

All tools are also available as CLI subcommands for shell scripts, hooks, and terminal workflows:

Common options: --json (JSON output), --compact (compact output), --limit N, --depth N, --file <path>.

As of v0.37.0 the CLI is clap-based: every subcommand has --help for its full flag list (code-graph-mcp <command> --help), value flags accept both --flag value and --flag=value, and unknown flags or malformed arguments fail fast with a clear error and a non-zero exit code (2) instead of being silently ignored. For example, trace hides downstream middleware with --no-middleware (shown by default), and snapshot is a create/inspect subcommand pair.

Plugin Skills

Installing the plugin ships two skills that Claude loads on its own when the situation matches — there are no slash commands to remember:

Both are thin routers over the CLI subcommands documented above, so anything a skill does is also runnable by hand.

Supported Languages (19)

Known limitations:

• Rust has no inherits edges — the language has no class inheritance, so impl Trait for Type is recorded as implements and an inherits-filtered query returns empty for Rust.

• Kotlin/Swift interface conformance is recorded as inherits (both use a single : Type grammar for base classes and protocols/interfaces), so implements-filtered queries return empty for these two languages.

• Cross-file dead-code detection may false-positive a type whose only cross-file reference sits beyond the 4096-byte stored-content cap per node (documented accepted limitation, v0.97.1).

Team-shared graph snapshot

Skip the full local index for team members and CI runners by publishing a ~3-5MB graph snapshot with each GitHub release.

Setup (one-time):

1. Copy node_modules/@sdsrs/code-graph/claude-plugin/templates/code-graph-snapshot.yml into your repo's .github/workflows/.

2. Push a release tag. The workflow uploads code-graph-snapshot-<sha>.db.zst as a release asset.

Verify:

npx -y -p @sdsrs/code-graph code-graph-mcp snapshot inspect ./code-graph-snapshot-<sha>.db.zst

After setup, the auto-fetch is opt-in per consumer: an untrusted repo could otherwise seed a misleading graph, so an unconfigured clone prints a hint and skips the install. Enable it with one of the trust signals below. These live in the environment (never in .code-graph.toml) so a committed/PR-injected config file cannot set them.

Integrity fail-closes: with no pin set and no fetchable <url>.blake3 sidecar, install is refused rather than accepting unverified content. Set a pin, or have the publisher serve the .blake3 sidecar alongside the snapshot. (file:// sources, used in tests, are exempt.)

Offline / air-gapped usage

The optional embedding model (needed for vector semantic search) is downloaded lazily on first use. To control that in a restricted environment:

Installing the model by hand

CODE_GRAPH_MODEL_DIR is the supported manual route. Hand-populating the default cache dir (<platform cache>/code-graph/models) does not work: that dir is only trusted once extract_and_promote has written a .model-id marker, so a manually filled copy is treated as not-current and re-downloaded.

# Download + verify against the published checksum, then extract flat
# (the archive has no wrapping folder: model.safetensors, tokenizer.json, config.json).
curl -sL -o models.tar.gz \
  https://github.com/sdsrss/code-graph-mcp/releases/download/v<version>/models.tar.gz
curl -sL https://github.com/sdsrss/code-graph-mcp/releases/download/v<version>/models.tar.gz.sha256
sha256sum -c models.tar.gz.sha256
mkdir -p ~/.cache/code-graph/models-manual
tar -xzf models.tar.gz -C ~/.cache/code-graph/models-manual
export CODE_GRAPH_MODEL_DIR=~/.cache/code-graph/models-manual

On Windows use %LOCALAPPDATA%\code-graph\models-manual, and give tar a POSIX path for -C (/c/Users/<you>/...) — tar -C C:\... fails with "Cannot connect to C: resolve failed". The model is pinned by content hash, so a manual install must be redone on each version bump.

Diagnosing a model that never downloads

code-graph-mcp doctor reports the model state, distinguishing "model files present but not loaded" (weights on disk — e.g. installed by the npm plugin — just no MCP server session yet) from "no download has been attempted" and from "download FAILED after N attempt(s): …" with the underlying error. The raw record is <platform cache>/code-graph/model-download.json. On a network that performs TLS inspection, the download retries automatically against the OS certificate store after the bundled roots are rejected.

Storage

Uses SQLite with:

• FTS5 for full-text search

• sqlite-vec extension for vector similarity search

• Merkle tree hashes for incremental change detection

Data is stored in .code-graph/index.db under the project root (auto-created, gitignored).

Environment variables

Every CODE_GRAPH_* variable the code reads, in one place — the alternative was reading the source to find out a switch existed. All of them are optional; the defaults are what you get by doing nothing.

Switches you may actually want

Build from Source

Prerequisites

• Rust 1.95.0 (2021 edition) — the toolchain CI and the release build pin (dtolnay/rust-toolchain@1.95.0 in .github/workflows/). No older toolchain is tested: Cargo.lock is lockfile version 4, which Cargo 1.75 cannot read at all, and no rust-version floor is declared in Cargo.toml.

• A C compiler (for bundled SQLite / sqlite-vec)

Build

# Default build — FTS5-only (~10 MB binary, no embedding model)
cargo build --release

# With local embedding model (~150 MB binary; downloads ~120 MB model lazily on first semantic search)
cargo build --release --features embed-model

Direct cargo install users get the FTS5-only build by default. npm/npx/plugin users get the full hybrid (FTS5 + vector) build automatically — release CI compiles with --features embed-model for shipped binaries.

Configure (from source)

Add the compiled binary to your MCP settings:

{
  "mcpServers": {
    "code-graph": {
      "command": "/path/to/target/release/code-graph-mcp"
    }
  }
}

Development

# Run tests
cargo test

# Run tests without embedding model
cargo test --no-default-features

# Check compilation
cargo check

# Run performance benchmarks (indexing, search, call graph)
cargo bench --no-default-features

License

See LICENSE for details.