Miru

What miru MCP server does

The miru MCP server helps coding agents find relevant code by combining semantic and keyword search. Results are returned as chunks that identify the file path, line range, and matching snippet. This makes it suitable for questions such as locating authentication middleware, configuration, or an implementation whose exact name is unknown.

The server is designed to reduce the amount of repository text an agent must inspect. It can replace broad grep or glob searches for code exploration, while still providing exact-match and context tools for follow-up work.

How it works

Agents normally start with search, which is the default tool for code exploration and accepts a question in natural language. Use one search call per question. When the required text is known exactly, locate searches for substrings such as environment variable names, symbols, or error codes.

A search result can indicate that its snippet is truncated. In that case, expand retrieves more context from the same file. find_related takes a file path and anchor line, then looks for similar code in other files. The separate read_benchmark tool reports cumulative token-savings data when benchmark mode is enabled; it is not part of the normal search workflow.

Setup and configuration

The miru MCP server can be installed globally with Bun using bun add -g @takara-ai/miru-code. Run miru setup to authenticate interactively, or select device-code login explicitly with miru setup --device. A bearer token can be stored with miru setup --key YOUR_TOKEN, and saved credentials can be removed with miru setup --clear.

Credentials are stored locally in a versioned credentials.json file and can be refreshed automatically. Setting TAKARA_API_KEY overrides stored credentials. The interactive MCP or plugin path can also initiate device-code authentication when credentials are missing.

Run miru install to select supported agent integrations from an interactive terminal interface. Depending on the selected client, installation can add the MCP configuration, instructions, rules, or optional skills. Restart the IDE after installation. Plugin credentials are separate from CLI credentials, so using both installation methods requires authentication in each scope.

Tools and capabilities

The miru MCP server exposes these core tools:

• search: Hybrid semantic and keyword search for general code exploration.
• locate: Exact substring lookup for known text, symbols, variables, and error codes.
• expand: Additional context from the same file when a result is truncated.
• find_related: Similar code in other files, starting from a file path and anchor line.
• read_benchmark: Cumulative token-savings reporting in benchmark mode.

The repository documents integrations for Claude Code, Cursor, Gemini CLI, Kiro, OpenCode, GitHub Copilot, Codex, VS Code, Visual Studio on Windows, and Windsurf or Devin Desktop. Plugin packages are documented for Claude Code, Codex, and Cursor, but their capabilities differ: the Cursor plugin currently provides skills and rules rather than an MCP entry, while Claude Code and Codex expose the listed MCP tools.

Limitations and notes

Plugin manifests launch the published runtime through bunx @takara-ai/miru-code@latest mcp. Local source changes therefore do not affect plugin behavior until a package version is published. A self-contained plugin installation that does not require Bun is not yet available.

Plugin installs do not provide identical features across IDEs. Dedicated sub-agent and benchmark configuration support are limited by the client plugin schemas, and benchmark mode is documented for the Claude Code plugin. Windsurf hooks are experimental and do not provide an MCP entry. Optional Caveman and STE features are skills rather than MCP search tools.