io.github.spuentesp/lain

Lain

A code graph for coding agents. Lain maps callers, dependencies, symbols, and Git co-changes, then exposes them through the Model Context Protocol (MCP). Before an agent edits a function, it can inspect what depends on it and how the rest of the code reaches it.

The demo traces a symbol across the bytes and tokio repositories. Agents query the same graph through MCP; the browser shows what Lain indexed and where an answer came from.

Ask your agent:

If I change validate_token, what breaks?
Show me the call chain from entry to helper_a.
Which files usually change with auth.rs?
Is another agent already editing this code?

Lain answers with symbol names and source paths from a local index. The index stays on disk between runs and can cover more than one repository. It follows commits, not uncommitted edits: after committing, a re-index pass (lain reindex or a fresh session) brings it current — and answers that may miss recent work say so instead of going quiet. Semantic search is optional; graph queries work without an embedding model.

Install

Using the release installer:

curl -fsSL https://raw.githubusercontent.com/spuentesp/lain/main/install.sh | bash
source ~/.zshrc   # or ~/.bashrc
lain --version

Or from npm (macOS, Linux, and x64 Windows):

npm install -g @spuentesp/lain-mcp
lain --version

Two npm notes:

• npm install -g fails with EEXIST when an unrelated lain file already sits in npm's global bin directory — for example an older manual copy in ~/.local/bin when your npm prefix is ~/.local. The npm error itself cannot tell the file is unrelated; move or remove it, or install under a different prefix: npm install -g --prefix ~/.local/lain-npm @spuentesp/lain-mcp (then add ~/.local/lain-npm/bin to your PATH).
• If npm install scripts were disabled (--ignore-scripts, a locked-down CI), the platform binary is not fetched during install. The first lain run downloads and verifies it, so that first run needs network access; once cached, lain works offline.

Rust 1.89 or newer is only needed when building from source. The quickstart covers non-interactive installation and the optional local embedding model.

Connect an agent

Run one setup command from the repository you want Lain to index:

lain setup --agent claude-code
lain setup --agent codex
lain setup --agent cursor
lain setup --agent vscode
lain setup --agent continue

Use lain setup --agent generic for another MCP host. Each adapter writes the client's own configuration — claude mcp add for Claude Code, ~/.codex/config.toml for Codex, ~/.cursor/mcp.json for Cursor, .vscode/mcp.json (or the user-scoped mcp.json) for VS Code, a block file under the repository's .continue/mcpServers/ for Continue. generic writes this project-level .mcp.json:

{
  "mcpServers": {
    "lain": {
      "command": "/home/you/.local/lain/lain",
      "args": ["mcp"]
    }
  }
}

command is always the absolute path of the lain binary that ran setup — not the bare name — so the agent runs the same binary no matter which PATH it inherits. That path depends on the install: the release installer puts the binary at ~/.local/lain/lain, while the npm package registers the verified cache binary it launches (~/.cache/lain/<version>/<target>/lain on Linux, ~/Library/Caches/lain/… on macOS, %LOCALAPPDATA%\lain\Cache\… on Windows). Setup finishes by starting exactly that command once and verifying the MCP handshake.

On a terminal, lain setup without --agent asks which client to configure (Enter picks Claude Code), and without --lsp asks which optional language servers to install (Enter installs none). Non-interactive runs default to generic and no language servers, so scripts should pass --agent and --lsp explicitly.

Restart the agent after setup. Lain walks up from the agent's working directory to find .git, builds .lain/graph.bin, and serves the repository over stdio. No repos.yaml is needed for one repository.

Languages

Parsers for Rust, Python, TypeScript, JavaScript, Go, Java, C, C++, C#, Ruby, Swift, Kotlin, Scala and PHP are built into the binary, as are the <script> blocks of Vue and Svelte components. Every one gets definitions and a call graph with nothing else installed.

Language servers are optional and only add precision. lain setup lists the languages it found and asks which servers to install; pressing Enter installs none. Non-interactive runs install none unless you name them:

lain setup --lsp python,go      # just these
lain setup --lsp detected       # every missing one for this repository
lain setup --lsp none

First query

Ask the connected agent:

Use Lain to show the blast radius of validate_token.

Or call the same tool from a terminal:

lain oneshot get_blast_radius validate_token

Use a symbol from your repository in place of validate_token. If the graph isn't ready or an answer looks stale, run lain doctor.

Multiple repositories

lain repos add bytes https://github.com/tokio-rs/bytes.git
lain repos add tokio https://github.com/tokio-rs/tokio.git
lain workspaces create tokio-stack --members bytes,tokio
lain server --config ./repos.yaml --transport http --port 9999

Open http://localhost:9999 for the Command Center. The server listens on localhost only; set LAIN_BIND_ADDR=0.0.0.0 (with LAIN_API_KEYS=key1,…) to expose it on the network. Federation adds organization-wide search and cross-repository blast-radius queries. The federation guide covers configuration, source types, and failure states. After a federation schema upgrade, run lain reindex to rebuild the on-disk graph from the sources.

Check the graph

From a source checkout:

scripts/demo.sh --quick

The script creates a small repository with a known call graph, starts a real Lain server, and compares the answers with that graph. It exits non-zero when a check fails. A full run also records timings against this repository.

What Lain provides

Static analysis can't prove every dynamic call through a message bus, dependency injection container, or reflective router. explain_dispatch combines static callers with convention matches, runtime traces, and Git co-change evidence; an empty result becomes insufficient_evidence, not a claim that the edit is safe.

Documentation

Project status

License

MIT — Copyright (c) 2026 spuentesp