astllm-mcp

MCP server for efficient code indexing and symbol retrieval. Index GitHub repos or local folders once with tree-sitter AST parsing, then let AI agents retrieve only the specific symbols they need — instead of loading entire files.

Simple 1 file binary distribution for trivial deployments.

Cut code-reading token costs by up to 99%.

How it works

1. Index — fetch source files, parse ASTs with tree-sitter, store symbols with byte offsets

2. Explore — browse file trees and outlines without touching file content

3. Retrieve — fetch only the exact function/class/method you need via O(1) byte-offset seek

4. Savings — every response reports tokens saved vs loading raw files

The index is stored locally in ~/.code-index/ (configurable). Incremental re-indexing only re-parses changed files.

The server automatically indexes the working directory on startup (incremental, non-blocking). Optionally set ASTLLM_WATCH=1 to also watch for file changes and re-index automatically. The watcher skips noisy directories (node_modules, .git, dist, .next, etc.) by default — see Watch excludes below.

Related MCP server: token-savior

Supported languages

Python, JavaScript, TypeScript, TSX, Go, Rust, Java, PHP, Dart, C#, C, C++, Dart/Flutter, Swift

Installation

Option 1: Download a pre-built binary (recommended)

Download the binary for your platform from the GitHub Releases page:

# Example for Linux x86-64
curl -L https://github.com/tluyben/astllm-mcp/releases/latest/download/astllm-mcp-linux-x86 -o astllm-mcp
chmod +x astllm-mcp
./astllm-mcp   # runs as an MCP stdio server

No Node.js, no npm, no build tools required.

Option 2: Build from source

Requires Node.js 18+ and a C++20-capable compiler (for tree-sitter native bindings).

git clone https://github.com/tluyben/astllm-mcp
cd astllm-mcp
CXXFLAGS="-std=c++20" npm install --legacy-peer-deps
npm run build

Note on Node.js v22+: The CXXFLAGS="-std=c++20" flag is required because Node.js v22+ v8 headers mandate C++20. The --legacy-peer-deps flag is needed because tree-sitter grammar packages target slightly different tree-sitter core versions.

MCP client configuration

Claude Code

Option A — claude mcp add CLI (easiest):

# Pre-built binary, project-scoped (.mcp.json)
claude mcp add astllm /path/to/astllm-mcp-linux-x86 --scope project

# Pre-built binary, user-scoped (~/.claude.json)
claude mcp add astllm /path/to/astllm-mcp-linux-x86 --scope user

# From source (Node.js), project-scoped
claude mcp add astllm node --args /path/to/astllm-mcp/dist/index.js --scope project

Option B — manual JSON config:

Add to ~/.claude.json (global) or .mcp.json in your project root (project-scoped):

Pre-built binary:

{
  "mcpServers": {
    "astllm": {
      "command": "/path/to/astllm-mcp-linux-x86",
      "type": "stdio"
    }
  }
}

From source (Node.js):

{
  "mcpServers": {
    "astllm": {
      "command": "node",
      "args": ["/path/to/astllm-mcp/dist/index.js"],
      "type": "stdio"
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

Pre-built binary:

{
  "mcpServers": {
    "astllm": {
      "command": "/path/to/astllm-mcp-macosx-arm"
    }
  }
}

From source (Node.js):

{
  "mcpServers": {
    "astllm": {
      "command": "node",
      "args": ["/path/to/astllm-mcp/dist/index.js"]
    }
  }
}

Tools

Indexing

index_repo

Index a GitHub repository. Fetches source files via the GitHub API, parses ASTs, stores symbols locally.

repo_url            GitHub URL or "owner/repo" slug
generate_summaries  Generate one-line AI summaries (requires API key, default: false)
incremental         Only re-index changed files (default: true)
storage_path        Custom storage directory

index_folder

Index a local folder recursively.

folder_path             Path to index
generate_summaries      AI summaries (default: false)
extra_ignore_patterns   Additional gitignore-style patterns
follow_symlinks         Follow symlinks (default: false)
incremental             Only re-index changed files (default: true)
storage_path            Custom storage directory

Navigation

list_repos

List all indexed repositories with file count, symbol count, and last-indexed time.

get_repo_outline

High-level overview: directory breakdown, language distribution, symbol kind counts.

repo    Repository identifier ("owner/repo" or short name if unique)

get_file_tree

File and directory structure with per-file language and symbol count. Much cheaper than reading files.

repo             Repository identifier
path_prefix      Filter to a subdirectory
include_summaries  Include per-file summaries

get_file_outline

All symbols in a file as a hierarchical tree (methods nested under their class).

repo       Repository identifier
file_path  File path relative to repo root

Retrieval

get_symbol

Full source code for a single symbol, retrieved by byte-offset seek (O(1)).

repo          Repository identifier
symbol_id     Symbol ID from get_file_outline or search_symbols
verify        Check content hash for drift detection (default: false)
context_lines Lines of context around the symbol (0–50, default: 0)

get_symbols

Batch retrieval of multiple symbols in one call.

repo        Repository identifier
symbol_ids  Array of symbol IDs

Search

search_symbols

Search symbols by name, kind, language, or file pattern. Returns signatures and summaries — no source loaded until you call get_symbol.

repo          Repository identifier
query         Search query
kind          Filter: function | class | method | type | constant | interface
file_pattern  Glob pattern, e.g. "src/**/*.ts"
language      Filter by language
limit         Max results 1–100 (default: 50)

search_text

Full-text search across indexed file contents. Useful for string literals, comments, config values.

repo          Repository identifier
query         Case-insensitive substring
file_pattern  Glob pattern to restrict files
limit         Max matching lines (default: 100)

Cache

invalidate_cache

Delete a repository's index, forcing full re-index on next operation.

repo    Repository identifier

Symbol IDs

Symbol IDs have the format file/path::qualified.Name#kind, for example:

src/auth/login.ts::AuthService.login#method
src/utils.go::parseURL#function
lib/models.py::User#class

Get IDs from get_file_outline or search_symbols, then pass them to get_symbol.

Token savings

Every response includes a _meta envelope:

{
  "_meta": {
    "timing_ms": 2.1,
    "tokens_saved": 14823,
    "total_tokens_saved": 89412,
    "cost_avoided_claude_usd": 0.222345,
    "cost_avoided_gpt_usd": 0.148230,
    "total_cost_avoided_claude_usd": 1.34118
  }
}

AI summaries (optional)

Set one of these environment variables to enable one-line symbol summaries:

# Anthropic Claude Haiku (recommended)
export ANTHROPIC_API_KEY=sk-ant-...

# Google Gemini Flash
export GOOGLE_API_KEY=...

# OpenAI-compatible (Ollama, etc.)
export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_MODEL=llama3

Summaries use a three-tier fallback: docstring first-line → AI → signature.

Environment variables

Legacy JASTLLM_* variable names are also accepted for compatibility with the original Python version's indexes.

Watch excludes

When ASTLLM_WATCH=1, the watcher walks the directory tree selectively — it opens one inotify watch per non-excluded directory, not per file, and skips the following by default:

node_modules  .git       dist        .next     .nuxt
build         out        __pycache__ .cache    target
vendor        venv       .venv       .tox      coverage
.nyc_output   .gradle    .idea       .vscode   .DS_Store
eggs          .mypy_cache .pytest_cache .ruff_cache

All hidden directories (.foo) are also skipped, except .github.

To add custom excludes, create ~/.astllm/exclude — one name per line, # for comments:

# ~/.astllm/exclude
my_large_assets_dir
some_vendor_folder
generated

Each name is matched against directory basenames anywhere in the tree, so generated excludes src/generated, lib/generated, etc.

Telling Claude to use this MCP

By default Claude will use Grep/Glob/Read to explore code. To make it prefer the MCP tools, add the following to your project's CLAUDE.md:

## Code search

An astllm-mcp index is available for this project. Prefer MCP tools over Grep/Glob/Read for all code exploration:

- `search_symbols` — find functions, classes, methods by name (use this first)
- `get_file_outline` — list all symbols in a file before deciding to read it
- `get_repo_outline` — understand project structure without reading files
- `get_symbol` — read a specific function/class source (O(1), much cheaper than reading the file)
- `get_symbols` — batch-read multiple symbols in one call
- `search_text` — full-text search for strings, comments, config values
- `get_file_tree` — browse directory structure with symbol counts

Only fall back to Grep/Read when the MCP tools cannot cover the case (e.g. a file type not indexed by tree-sitter).

The repo identifier to pass to MCP tools is local/<folder-name> for locally indexed folders (e.g. local/src). Use list_repos if unsure.

Security

• Path traversal and symlink-escape protection

• Secret files excluded (.env, *.pem, *.key, credentials, etc.)

• Binary files excluded by extension and content sniffing

• File size limits enforced before reading

Single-file binaries (no Node.js required)

Uses Bun to produce self-contained executables. All JS and native tree-sitter .node addons are embedded — users just download and run, no npm install or Node.js needed.

Prerequisites: install Bun once (curl -fsSL https://bun.sh/install | bash), then:

npm run build:macosx-arm   # → dist/astllm-mcp-macosx-arm  (run on macOS ARM)
npm run build:linux-x86    # → dist/astllm-mcp-linux-x86   (run on Linux x86)
npm run build:linux-arm    # → dist/astllm-mcp-linux-arm   (run on Linux ARM)

Each build script must run on the matching platform. The grammar packages ship prebuilt .node files for all platforms, but the tree-sitter core is compiled from source on install. scripts/prep-bun-build.mjs (run automatically before each binary build) copies the compiled .node into the location Bun expects. For CI, use a matrix — Linux x86 and Linux ARM can both build on Linux via Docker/QEMU; macOS ARM requires a macOS runner.

How it works: tree-sitter and all grammar packages support bun build --compile via a statically-analyzable require() path. Bun embeds the correct native addon for the target and extracts it to a temp directory on first run.

Development

npm run build   # compile TypeScript → dist/
npm run dev     # run directly with tsx (no compile step)

The project is TypeScript ESM. All local imports use .js extensions (TypeScript NodeNext resolution).

Storage layout

~/.code-index/
  <owner>/
    <repo>/
      index.json        # symbol index with byte offsets
      files/            # raw file copies for byte-offset seeking
        src/
          auth.ts
          ...
  _savings.json         # cumulative token savings

Inspiration

This tool was inspired by: https://github.com/jgravelle/jcodemunch-mcp

I needed simplified distribution and a bunch of features this did not have.