Kin

What Kin MCP server does

The Kin MCP server gives compatible AI coding clients access to a graph-native representation of a code repository. Kin records source, functions, types, and relationships between entities as repository state, while preserving exact source and maintaining versioned history. The graph is stored locally outside the working checkout, so the repository itself is not modified during initialization.

The server is useful for questions such as where a symbol is defined, which code refers to it, what nearby context surrounds it, and what parts of the graph may be affected by a change. These queries are based on recorded relationships rather than only matching text. Kin also supports lexical retrieval without a model; embeddings and reranking are optional.

How it works

Initialize a full clone or use kin clone to create a Kin-aware checkout. During initialization, Kin reads Git history and admits repository content into its local store. By default, the checked-out commit becomes available first, while reachable history can be imported in the background. The store normally lives under ~/.kin/stores, and kin store path reports the location for a checkout.

The Kin MCP server and command-line interface query the same graph. Typical CLI checks include kin locate for finding an entity, kin refs for recorded references, kin trace for nearby context, and kin impact for graph-based effect exploration. Agents can use the corresponding repository knowledge through MCP clients, while developers can compare the results with the source code.

For semantic retrieval, an operator can configure an OpenAI-compatible embedding endpoint with environment variables such as KIN_MODEL_EMBEDDINGS_PROVIDER, KIN_MODEL_EMBEDDINGS_BASE_URL, and KIN_MODEL_EMBEDDINGS_MODEL. Graph and lexical queries do not require downloading a model.

Setup and configuration

Install on macOS, Linux, or WSL2 with Node.js 20 or newer using npx -y @kinlab/kin setup, or use the shell installer. The setup command installs the CLI and can connect AI coding tools it detects. Run kin setup again after adding another client; kin setup status reports configured connections.

After installation, run kin init . inside an existing repository, or start with kin clone <url>. kin overview, kin status, and kin graph status help confirm imported entities, repository state, and graph coverage. Claude Code, Codex, Cursor, Gemini, and VS Code are named integration targets in the project documentation.

Limitations and notes

Kin is in public beta, so behavior may change and rough edges are expected. A shallow clone admits only its checked-out commit. Submodules and Git LFS are unsupported, and Git hooks, sparse checkouts, or unfinished merges and rebases can prevent import.

Imported history consumes time, memory, and disk according to the amount of history. Uncommitted and untracked files are excluded from imported Git history. Native Windows installation is available for x64, but it does not connect AI coding tools automatically; WSL2 is the recommended Windows path. No native Windows ARM64 build is published.

The Kin MCP server does not make review decisions or block merges. Kin’s review report is advisory, and its authorship information is declared rather than independently verified.