# mcp-codebase-index [](https://pypi.org/project/mcp-codebase-index/) [](https://github.com/MikeRecognex/mcp-codebase-index/actions/workflows/ci.yml) [](https://www.python.org/downloads/) [](https://www.gnu.org/licenses/agpl-3.0) [](https://modelcontextprotocol.io) []() A structural codebase indexer with an [MCP](https://modelcontextprotocol.io) server for AI-assisted development. Zero runtime dependencies — uses Python's `ast` module for Python analysis and regex-based parsing for TypeScript/JS, Go, Rust, and C#. Requires Python 3.11+. ## What It Does Indexes codebases by parsing source files into structural metadata -- functions, classes, imports, dependency graphs, and cross-file call chains -- then exposes 18 query tools via the Model Context Protocol, enabling Claude Code and other MCP clients to navigate codebases efficiently without reading entire files. **Automatic incremental re-indexing:** In git repositories, the index stays up to date automatically. Before every query, the server checks `git diff` and `git status` (~1-2ms). If files changed, only those files are re-parsed and the dependency graph is rebuilt. No need to manually call `reindex` after edits, branch switches, or pulls. **Persistent disk cache:** The index is saved to a pickle cache file (`.codebase-index-cache.pkl`) after every build. On subsequent server starts, the cache is loaded and validated against the current git HEAD — if the ref matches, startup is instant. If a small number of files changed (≤20), the cached index is loaded and incrementally updated instead of rebuilt from scratch. This eliminates the cold-start penalty when restarting Claude Code sessions, restarting the MCP server, or resuming work after context compaction. ## Language Support | Language | Method | Extracts | |----------|--------|----------| | Python (`.py`) | AST parsing | Functions, classes, methods, imports, dependency graph | | TypeScript/JS (`.ts`, `.tsx`, `.js`, `.jsx`) | Regex-based | Functions, arrow functions, classes, interfaces, type aliases, imports | | Go (`.go`) | Regex-based | Functions, methods (receiver-based), structs, interfaces, type aliases, imports, doc comments | | Rust (`.rs`) | Regex-based | Functions (`pub`/`async`/`const`/`unsafe`), structs, enums, traits, impl blocks, use statements, attributes, doc comments, macro_rules | | C# (`.cs`) | Regex-based | Classes, interfaces, structs, enums, records, methods, constructors, using directives, `[Attributes]`, `///` XML doc comments | | Markdown/Text (`.md`, `.txt`, `.rst`) | Heading detection | Sections (# headings, underlines, numbered, ALL-CAPS) | | Other | Generic | Line counts only | ## Installation ```bash pip install "mcp-codebase-index[mcp]" ``` The `[mcp]` extra includes the MCP server dependency. Omit it if you only need the programmatic API. For development (from a local clone): ```bash pip install -e ".[dev,mcp]" ``` ## MCP Server ### Running ```bash # As a console script PROJECT_ROOT=/path/to/project mcp-codebase-index # As a Python module PROJECT_ROOT=/path/to/project python -m mcp_codebase_index.server ``` `PROJECT_ROOT` specifies which directory to index. Defaults to the current working directory. ### Persistent Cache In git repositories, the server automatically caches the index to `.codebase-index-cache.pkl` in the project root. On startup: 1. **Cache hit (exact match):** If the cached git ref matches the current HEAD, the index loads instantly from disk — no parsing, no file walking. 2. **Cache hit (small changeset):** If ≤20 files changed since the cached ref, the cached index is loaded and incrementally updated on the first query. 3. **Cache miss:** If the changeset is large or no cache exists, a full rebuild runs and saves a new cache. Add `.codebase-index-cache.pkl` to your `.gitignore` — it's a local-only build artifact. ### Configuring with OpenClaw Install the package on the machine where OpenClaw is running: ```bash # Local install pip install "mcp-codebase-index[mcp]" # Or inside a Docker container / remote VPS docker exec -it openclaw bash pip install "mcp-codebase-index[mcp]" ``` Add the MCP server to your OpenClaw agent config (`openclaw.json`): ```json { "agents": { "list": [{ "id": "main", "mcp": { "servers": [ { "name": "codebase-index", "command": "mcp-codebase-index", "env": { "PROJECT_ROOT": "/path/to/project" } } ] } }] } } ``` Restart OpenClaw and verify the connection: ```bash openclaw mcp list ``` All 18 tools will be available to your agent. **Performance note:** The server automatically detects file changes via `git diff` before every query (~1-2ms) and incrementally re-indexes only what changed. However, OpenClaw's default MCP integration via mcporter spawns a fresh server process per tool call, which discards the in-memory index and forces a full rebuild each time (~1-2s for small projects, longer for large ones). With persistent caching, these cold starts are now significantly faster — the server loads from the disk cache instead of re-parsing the entire codebase. For persistent connections (avoiding even the cache load overhead), use the [openclaw-mcp-adapter](https://github.com/androidStern-personal/openclaw-mcp-adapter) plugin, which connects once at startup and keeps the server running: ```bash pip install openclaw-mcp-adapter ``` ### Configuring with Claude Code Add to your project's `.mcp.json`: ```json { "mcpServers": { "codebase-index": { "command": "mcp-codebase-index", "env": { "PROJECT_ROOT": "/path/to/project" } } } } ``` Or using the Python module directly (useful if installed in a virtualenv): ```json { "mcpServers": { "codebase-index": { "command": "/path/to/.venv/bin/python3", "args": ["-m", "mcp_codebase_index.server"], "env": { "PROJECT_ROOT": "/path/to/project" } } } } ``` #### Reinforcing Tool Usage with Hooks Claude Code tends to default to built-in Glob/Grep/Read tools even when codebase-index is available. In addition to CLAUDE.md instructions (see below), you can add hooks that fire on every prompt to reinforce the behavior. Add this to `.claude/settings.local.json`: ```json { "hooks": { "SessionStart": [ { "hooks": [ { "type": "command", "command": "echo 'CRITICAL REMINDER: Use codebase-index MCP tools FIRST for ALL code navigation (find_symbol, get_function_source, search_codebase, get_dependencies, etc). Only fall back to Glob/Grep/Read for non-code files.'" } ] } ], "UserPromptSubmit": [ { "hooks": [ { "type": "command", "command": "echo 'Use codebase-index MCP tools first for code navigation.'" } ] } ] } } ``` Hook stdout is injected as context Claude sees before responding. `SessionStart` fires on startup, resume, and context compaction. `UserPromptSubmit` fires on every turn. ### Important: Make the AI Actually Use Indexed Tools By default, AI assistants will ignore the indexed tools and fall back to reading entire files with Glob/Grep/Read. Soft language like "prefer" gets rationalized away. Add this to your project's `CLAUDE.md` (or equivalent instructions file) with **mandatory** language: ``` ## Codebase Navigation — MANDATORY You MUST use codebase-index MCP tools FIRST when exploring or navigating the codebase. This is not optional. - ALWAYS start with: get_project_summary, find_symbol, get_function_source, get_class_source, get_structure_summary, get_dependencies, get_dependents, get_change_impact, get_call_chain, search_codebase - Only fall back to Read/Glob/Grep when codebase-index tools genuinely don't have what you need (e.g. reading non-code files, config, frontmatter) - If you catch yourself reaching for Glob/Grep/Read to find or understand code, STOP and use codebase-index instead ``` The word "prefer" is too weak — models treat it as a suggestion and default to familiar tools. Mandatory language with explicit fallback criteria is what actually changes behavior. ### Available Tools (18) | Tool | Description | |------|-------------| | `get_project_summary` | File count, packages, top classes/functions | | `list_files` | List indexed files with optional glob filter | | `get_structure_summary` | Structure of a file or the whole project | | `get_functions` | List functions with name, lines, params | | `get_classes` | List classes with name, lines, methods, bases | | `get_imports` | List imports with module, names, line | | `get_function_source` | Full source of a function/method | | `get_class_source` | Full source of a class | | `find_symbol` | Find where a symbol is defined (file, line, type) | | `get_dependencies` | What a symbol calls/uses | | `get_dependents` | What calls/uses a symbol | | `get_change_impact` | Direct + transitive dependents | | `get_call_chain` | Shortest dependency path (BFS) | | `get_file_dependencies` | Files imported by a given file | | `get_file_dependents` | Files that import from a given file | | `search_codebase` | Regex search across all files (max 100 results) | | `reindex` | Force full re-index (rarely needed — incremental updates happen automatically in git repos) | | `get_usage_stats` | Session efficiency stats: tool calls, characters returned vs total source, estimated token savings | ## Benchmarks Tested across four real-world projects on an M-series MacBook Pro, from a small project to CPython itself (1.1 million lines): ### Index Build Performance | Project | Files | Lines | Functions | Classes | Index Time | Peak Memory | |---------|------:|------:|----------:|--------:|-----------:|------------:| | RMLPlus | 36 | 7,762 | 237 | 55 | 0.9s | 2.4 MB | | FastAPI | 2,556 | 332,160 | 4,139 | 617 | 5.7s | 55 MB | | Django | 3,714 | 707,493 | 29,995 | 7,371 | 36.2s | 126 MB | | **CPython** | **2,464** | **1,115,334** | **59,620** | **9,037** | **55.9s** | **197 MB** | With persistent caching, subsequent startups bypass the full build entirely. Cache load time is negligible compared to parsing — a cache hit on CPython restores the full index in under a second instead of 56s. ### Query Response Size vs Total Source Querying CPython — 41 million characters of source code: | Query | Response | Total Source | Reduction | |-------|-------:|------------:|----------:| | `find_symbol("TestCase")` | 67 chars | 41,077,561 chars | **99.9998%** | | `get_dependencies("compile")` | 115 chars | 41,077,561 chars | **99.9997%** | | `get_change_impact("TestCase")` | 16,812 chars | 41,077,561 chars | **99.96%** | | `get_function_source("compile")` | 4,531 chars | 41,077,561 chars | **99.99%** | | `get_function_source("run_unittest")` | 439 chars | 41,077,561 chars | **99.999%** | `find_symbol` returns 54-67 characters regardless of whether the project is 7K lines or 1.1M lines. Response size scales with the answer, not the codebase. `get_change_impact("TestCase")` on CPython found **154 direct dependents and 492 transitive dependents** in 0.45ms — the kind of query that's impossible without a dependency graph. Use `max_direct` and `max_transitive` to cap output to your token budget. ### Query Response Time All targeted queries return in sub-millisecond time, even on CPython's 1.1M lines: | Query | RMLPlus | FastAPI | Django | CPython | |-------|--------:|--------:|-------:|--------:| | `find_symbol` | 0.01ms | 0.01ms | 0.03ms | 0.08ms | | `get_dependencies` | 0.00ms | 0.00ms | 0.00ms | 0.01ms | | `get_change_impact` | 0.02ms | 0.00ms | 2.81ms | 0.45ms | | `get_function_source` | 0.01ms | 0.02ms | 0.03ms | 0.10ms | Run the benchmarks yourself: `python benchmarks/benchmark.py` ## How Is This Different from LSP? LSP answers "where is this function?" — mcp-codebase-index answers "what happens if I change it?" LSP is point queries: one symbol, one file, one position. It can tell you where `LLMClient` is defined and who references it. But ask "what breaks transitively if I refactor `LLMClient`?" and LSP has nothing. This tool returns 11 direct dependents and 31 transitive impacts in a single call — 204 characters. To get the same answer from LSP, the AI would need to chain dozens of find-reference calls recursively, reading files at every step, burning thousands of tokens to reconstruct what the dependency graph already knows. LSP also requires you to install a separate language server for every language in your project — pyright for Python, vtsls for TypeScript, gopls for Go. Each one is a heavyweight binary with its own dependencies and configuration. mcp-codebase-index is zero dependencies, handles Python + TypeScript/JS + Go + Rust + C# + Markdown out of the box, and every response has built-in token budget controls (`max_results`, `max_lines`). LSP was built for IDEs. This was built for AI. ## Programmatic Usage ```python from mcp_codebase_index.project_indexer import ProjectIndexer from mcp_codebase_index.query_api import create_project_query_functions indexer = ProjectIndexer("/path/to/project", include_patterns=["**/*.py"]) index = indexer.index() query_funcs = create_project_query_functions(index) # Use query functions print(query_funcs["get_project_summary"]()) print(query_funcs["find_symbol"]("MyClass")) print(query_funcs["get_change_impact"]("some_function")) ``` ## Development ```bash pip install -e ".[dev,mcp]" pytest tests/ -v ruff check src/ tests/ ``` ## References The structural indexer was originally developed as part of the [RMLPlus](https://github.com/MikeRecognex/RMLPlus) project, an implementation of the [Recursive Language Models](https://arxiv.org/abs/2512.24601) framework. ## License This project is dual-licensed: - **AGPL-3.0** for open-source use — see [LICENSE](LICENSE) - **Commercial License** for proprietary use — see [COMMERCIAL-LICENSE.md](COMMERCIAL-LICENSE.md) If you're using mcp-codebase-index as a standalone MCP server for development, the AGPL-3.0 license applies at no cost. If you're embedding it in a proprietary product or offering it as part of a hosted service, you'll need a commercial license. See [COMMERCIAL-LICENSE.md](COMMERCIAL-LICENSE.md) for details.