libsrc

An MCP server that resolves project dependencies and provides local source code paths for AI agent inspection. It auto-detects the build system, resolves dependencies, looks up source repositories, clones them, and checks out the correct version as a git worktree -- all in a single tool call.

Why

AI coding agents work best when they can inspect the source code of libraries used in a project. But getting at those sources is harder than it should be:

• Published packages are opaque -- Java distributes compiled JARs, and even when a -sources.jar is available, it has to be fetched separately and extracted. Python wheels contain bytecode and stripped metadata. In both cases, getting readable source into the agent's hands is needlessly hard.

• Packages lose context -- Original source repositories often contain documentation, examples, and markdown files that are stripped from published packages.

• Exact versions matter -- Agents need the precise version a project depends on, not "latest" or "close enough". libsrc resolves the exact version from lock files and build tools, then checks out the matching git tag.

• Standard tools just work on cloned repos -- Giving an agent a local path to a git checkout lets it explore with standard file tools, easily and predictably. No JAR extraction, no archive unpacking, no guessing -- just a normal directory of source files.

• Parallel-friendly -- Each version gets its own git worktree, so multiple agents (or the same agent across tasks) can inspect different versions of the same library concurrently without conflicts.

libsrc bridges this gap: one tool call turns a dependency name into a local path the agent can explore immediately.

Related MCP server: Cross-Project MCP Server

Supported Ecosystems

• Java: Maven (pom.xml), Gradle (build.gradle, build.gradle.kts)

• Python: Poetry (pyproject.toml + poetry.lock), uv (pyproject.toml + uv.lock)

Plain pip / requirements.txt is not supported (no reliable version resolution without a lock file).

Installation

# Install as a tool
uv tool install libsrc-mcp

# Or run directly without installing
uvx libsrc-mcp serve

# Auto-register in detected AI coding tools
libsrc-mcp install

Auto-install into AI tools

libsrc-mcp install detects installed AI coding tools and adds the MCP server to their configs. Supported: Claude Code, Cursor, Windsurf, Codex CLI, Gemini CLI, JetBrains Junie, VS Code Copilot. Skips tools that aren't installed or already configured.

libsrc-mcp install
libsrc-mcp install --port 8080  # if using a non-default port

Configuration

Create ~/.config/libsrc/config.yml (all fields optional):

# Directory where library sources are cloned (default: ~/devel/libs/)
output_dir: ~/devel/libs/

# HTTP server port (default: 7890)
port: 7890

# Trusted git hostings for cloning (default: github.com, gitlab.com)
trusted_hosts:
  - github.com
  - gitlab.com

# deps.dev API cache TTL in hours (default: 24)
deps_dev_cache_ttl: 24

Usage

Start the server

libsrc-mcp serve
libsrc-mcp serve --port 8080

Systemd service (Linux)

mkdir -p ~/.config/systemd/user
cp systemd/libsrc-mcp.service ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now libsrc-mcp

Worktree cleanup

Worktrees not accessed for 7+ days are cleaned up automatically on server startup. Manual cleanup:

libsrc-mcp cleanup

MCP Tool: get_library_sources

Without library_name: returns the dependency listing for discovery. With library_name: resolves source repos, clones, creates version worktrees, returns local paths.

How It Works

Dependency Resolution

Lock files are preferred when available, CLI tools are the fallback:

Build tool wrappers are preferred: mvnw > mvn, gradlew > gradle.

Source Repository Discovery

Layered approach (first match wins):

1. deps.dev API -- Google's unified package-to-repo mapping (Maven, PyPI, npm, Go, etc.). Responses are cached locally.

2. Registry fallback -- Maven Central POM <scm> element; PyPI JSON API project_urls.

3. Heuristics -- Maven groupId patterns (com.github.*), package name matching.

Clone and Worktree Management

• Full clones into <output_dir>/<hostname>/<owner>/<repo>

• git fetch --all --tags before each tag lookup

• Version worktrees at <clone_path>.versions/<version> (enables parallel agents on different versions)

• Tag matching: tries v{ver}, {ver}, release-{ver}, {artifact}-{ver}, suffix/contains fallback

• Strips release qualifiers (.Final, .RELEASE, .GA) for projects that omit them from tags

• Monorepo dedup: same repo+version = one worktree shared across artifacts

• File-based locking (fcntl.flock) for concurrent safety

Caching

• Dependency cache: ~/.cache/libsrc/deps/ keyed by content hash of all build/lock files (invalidates on any file change)

• deps.dev cache: ~/.cache/libsrc/depsdev/ with configurable TTL

• Worktree tracker: ~/.cache/libsrc/worktree-access.json for LRU cleanup

MCP Client Configuration

Claude Code

claude mcp add libsrc --transport http http://127.0.0.1:7890/mcp

VS Code Copilot (.vscode/mcp.json)

{
  "servers": {
    "libsrc": { "type": "http", "url": "http://127.0.0.1:7890/mcp" }
  }
}

Other tools

{
  "mcpServers": {
    "libsrc": { "type": "http", "url": "http://127.0.0.1:7890/mcp" }
  }
}

See docs/mcp-auto-install.md for tool-specific config details (Cursor, Windsurf, Gemini CLI, Codex CLI, Cline, Continue.dev, JetBrains Junie).

Releasing

Version is derived automatically from git tags via hatch-vcs — no manual version bumping needed.

git tag v<version>
git push origin v<version>

The publish.yml GitHub Action builds and publishes to PyPI automatically via trusted publishing.