ruv-swarm MCP
The ruv-swarm package runs a stdio MCP server: 17 tools for SQLite task metadata, forecasts and local benchmarks.
- 393 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The package builds a native SQLite binding and writes task state to local files.
Install
- Works in
Manual install
npm install ruv-swarmPackage install from the README, then `npx ruv-swarm mcp start` as a stdio server.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The package builds a native SQLite binding and writes task state to local files.
This is third-party code. Review the repository files before installing.
What it does
ruv-swarm is a Node.js package for coordinating agents, keeping task state in SQLite and calling the WebAssembly modules of the ruv-FANN repository. It includes a stdio MCP server for clients such as Claude Code. Start it with `npx ruv-swarm mcp start`, and `ruv-swarm mcp tools` lists the tools. The server exposes 17 supported tools: persisted local metadata, dense neural inference, statistical forecasts, memory observations and measured local benchmarks. MCP training tools, autonomous DAA execution, learned-pattern analysis and external model execution are unavailable and return explicit errors. Dense backpropagation training lives separately in the package JavaScript API.
Who it is for. Developers experimenting with agent coordination and small WebAssembly neural networks who want to connect it to Claude Code.
Good fit when
- Agent task state is kept in a local SQLite database and read through MCP tools.
- The agent builds a statistical forecast over a value series and compares intervals.
- A small dense network runs on WebAssembly and its time is measured locally.
- Agent observations are written to memory and read in the next session.
Not a fit when
- Network training through MCP is needed: the training tools are unavailable and return an error.
- The machine has no C++20 toolchain or Python 3: the package builds its SQLite binding from source.
- A WebSocket server is needed: it is not implemented in this distribution.
Example request
List the ruv-swarm tools, build a forecast for this series and save an observation to memory.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
RUV_SWARM_DB_PATH- Database path for persistent state
Limitations
The package needs Node.js 22.13 or newer, Python 3 and a C++20 build toolchain: the SQLite binding is built from verified source and there is no prebuilt fallback when the build fails. The npm package and Rust crates are versioned separately, so pin the version in clients. A JSON-RPC frame is capped at 1 MiB and a process accepts up to 32 pending requests. Training tools, autonomous DAA execution and external models do not work through MCP. The project is large and changes quickly, and features from older descriptions may be missing.
How to disable. Remove the `ruv-swarm` entry from the client settings. SQLite database files stay on disk and are deleted by hand.
FAQ
How do I start the MCP server?
Run `npx ruv-swarm mcp start`. In the client set `command` to `npx` and `args` to `["ruv-swarm", "mcp", "start"]`.
What is needed to install?
Node.js 22.13 or newer, Python 3 and a C++20 build toolchain. Install the package with `npm install ruv-swarm`.
How many tools does the server have?
17 supported ones. `ruv-swarm mcp tools` prints the list.
Can I train a network through MCP?
No, training tools are unavailable. Backpropagation training exists only in the package JavaScript API.
Does WebSocket work?
No, the WebSocket server is not implemented in this distribution. Stdio remains.
Discussion
No questions or reviews yet.
Related
MCP serversThe ak command installs, repairs and checks ruflo and agentic-qe, adds the offline RuvNet Brain base and links Claude Code, Codex
A Claude Code and Codex plugin with a local RuvNet source base: the agent finds the right tool and shows the file with its code.
A CLI and MCP server that builds a local wiki with a knowledge graph from documents, code and links, so the agent searches it instead of scanning files
A single Rust gateway between an agent and all its MCP servers: the client sees 9 to 17 meta-tools and finds the rest on demand.
A bundle of 176 skills, subagents and an MCP server for Claude Code, Cursor, Codex and Windsurf with code checks
Plugin marketplace and CLI for Claude Code: agent swarms, background workers, vector memory, security review and test generation
README in short
The Node.js package README covers installation, starting the MCP server, frame and queue limits and training a small network through WebAssembly.