MCP

Codelith has already read your repository — chunked, embedded, with an import graph, pinned to a commit. Coding agents re-derive that by grepping, from scratch, every session. The MCP server lets them ask instead.

It is not an app. It adds nothing of its own; it is a second transport over the same base.

Setup

// .mcp.json — Claude Code, Cursor, and anything else speaking MCP
{
  "mcpServers": {
    "codelith": {
      "command": "codelith",
      "args": ["mcp"]
    }
  }
}

Or, in Claude Code:

claude mcp add codelith -- codelith mcp

The server speaks JSON-RPC on stdin and stdout, which is what an editor launches. It writes nothing to stdout but the protocol — a stray line of prose there is a parse error on the client, which is why codelith mcp prints no banner.

The eight tools

Call list_codebases first. Every other tool takes a codebase_id, and the ids are not guessable.

Tool	Answers
list_codebases	What has been analysed, with commit and size
before_edit	What an edit would touch. Call this first
search_code	Search the source, by meaning or by name
read_file	A file as the knowledge base holds it
find_callers	Who calls a symbol. No embedding encodes this
find_dependents	Which files import a file
blast_radius	Everything that transitively reaches a file, with distance
list_facts	Routes, datastores, env vars, entrypoints and the rest

before_edit

The one that changes what this is for. The others answer questions about a codebase; this one is called before changing it.

Give it a path or a symbol:

codelith/llm/client.py — high risk to edit — 3 file(s) import it directly,
11 reach it within 3 hops, nothing tests it, 2 written page(s) describe it.

Breaks first (nearest first):
  1 hop(s)  codelith/agents/base.py
  2 hops(s) codelith/workflows/analysis_workflow.py
  ...

No test file reaches this. A change here is unverified by the suite.

Written pages that describe it (they will need re-writing):
  reference/llm-client — The LLM client

No model call and four indexed queries. A check an agent skips under time pressure is a check that does not exist.

What it will not do

• Nothing writes. Every tool is read-only.
• Nothing leaves. No network call beyond the model endpoint you configured.
• Nothing is guessed. find_callers returning empty says so explicitly, because call edges are resolved only where the callee is unambiguous and an empty result is not proof of no callers.

When answers look stale

The knowledge base is pinned to a commit. If you have moved on since analysing, the tools are describing the code as it was — accurately, and not as it is now.

list_codebases reports the commit each reading is at. Re-analyse to catch up, and the old reading stays, which is what What changed compares.