RELATED 01
Ruflo
Agent swarms with shared vector memory and task routing for Claude Code and other MCP clients
81kGitHub stars
87k/wknpm downloads
89popularity score
ruv-swarm
Local multi-agent swarm coordination with small WebAssembly neural networks, run from npx
By rUv (individual) TypeScript Local · stdio AI and MLDeveloper tools
56 popularity score ★ 387 ↓ 3.1k/wk
Connect your client
Local (npm)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add ruv-swarm -- npx ruv-swarm mcp start --protocol=stdioClaude Desktopclaude_desktop_config.json
{
"mcpServers": {
"ruv-swarm": {
"command": "npx",
"args": [
"ruv-swarm",
"mcp",
"start",
"--protocol=stdio"
]
}
}
}Cursor~/.cursor/mcp.json
{
"mcpServers": {
"ruv-swarm": {
"command": "npx",
"args": [
"ruv-swarm",
"mcp",
"start",
"--protocol=stdio"
]
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"ruv-swarm": {
"type": "stdio",
"command": "npx",
"args": [
"ruv-swarm",
"mcp",
"start",
"--protocol=stdio"
]
}
}
}Codex~/.codex/config.toml
[mcp_servers.ruv-swarm]
command = "npx"
args = ["ruv-swarm", "mcp", "start", "--protocol=stdio"]Gemini CLITerminal
gemini mcp add --scope user ruv-swarm npx -- ruv-swarm mcp start --protocol=stdioSettings 4
LOG_LEVEL |
Optional: Log level for the server (INFO by default, DEBUG with --debug). |
|---|---|
MCP_STABILITY |
Optional: Set to true to run under the auto-restart wrapper, same as the --stability flag. |
LOG_TO_FILE |
Optional: Set to true to also write logs to files. |
LOG_DIR |
Optional: Folder for log files when LOG_TO_FILE is on (./logs by default). |
What it does
ruv-swarm is the MCP server in ruvnet's ruv-FANN repository. The repository also holds a Rust port of the FANN neural network library. The server runs on your machine as a Node.js process. It gives an assistant tools to set up a swarm with a chosen topology (mesh, hierarchical, ring or star), add agents with roles, hand a task to the swarm, check its progress and fetch results. It also reports agent and memory metrics, runs benchmarks, detects runtime features such as WASM SIMD, and trains or inspects the small neural models attached to agents. A second set of tools, prefixed daa_, covers its 'decentralized autonomous agents' layer: adaptive agents, workflows, knowledge sharing between agents and meta-learning.
The swarm is bookkeeping and neural-model state inside the server process and its local store. The agents do not call another AI model, and no tool edits your code or runs commands. The assistant still does the actual work. The tools hand it a plan, a status and metrics to coordinate with.
The tools/list handler in bin/ruv-swarm-secure.js, the program npx starts, registers 25 tools: 15 core tools plus the 10 daa_ tools. The npm README's tool table lists memory_store and memory_get, but neither is registered there; memory_usage only reports memory statistics. The same file also serves two documentation resources (swarm://docs/getting-started and swarm://docs/stability).
What it needs
| Runtime | Node.js 14+ |
|---|
Node.js 14 or later (the npm page recommends 18+), with WebAssembly support; the package ships pre-built WASM binaries. Started over stdio with `npx ruv-swarm mcp start`. It needs no account, API key or network service.
Limits
Only stdio is implemented: in bin/ruv-swarm-secure.js a protocol other than stdio is refused ('WebSocket protocol not yet implemented'). The README's `mcp start --port 3000`, its ws://localhost:3000/mcp example and its MCP_ENABLED Docker example therefore do not give a working network endpoint. swarm_init accepts at most 100 agents (default 5). The speed and SWE-Bench figures in the README are the makers' own claims. Its agents coordinate and keep state, but they do not write code or call other models themselves.
Who it suits
Suits Claude Code users who want to try the makers' swarm and agent-coordination workflow, or to experiment with small CPU-run neural models, all locally and with no keys. It is a poor fit if you want a server that acts on outside systems: it reaches no API, repository or database, and its task tools track work rather than do it. Anyone who needs a network-reachable MCP endpoint should know that only stdio works today.
Tools 25
swarm_init- Create a swarm with a topology, agent cap and distribution strategy
swarm_status- Show the swarm's state and its agents
swarm_monitor- Watch swarm activity for a set period
agent_spawn- Add an agent of a given type to the swarm
agent_list- List the swarm's agents, optionally filtered
agent_metrics- Report performance figures for agents
task_orchestrate- Hand a task to the swarm with a priority and strategy
task_status- Check progress of running tasks
task_results- Fetch the results of a finished task
benchmark_run- Run the built-in performance benchmarks
features_detect- Report runtime capabilities such as WASM and SIMD support
memory_usage- Report current memory use
neural_status- Show neural model status and metrics for agents
neural_train- Train agents' neural models for a number of iterations
neural_patterns- Describe the cognitive patterns agents can use
daa_init- Start the autonomous-agent layer
daa_agent_create- Create an adaptive autonomous agent
daa_agent_adapt- Feed feedback to an agent so it adjusts its behaviour
daa_workflow_create- Define a multi-step workflow for autonomous agents
daa_workflow_execute- Run a defined workflow on chosen agents
daa_knowledge_share- Pass knowledge from one agent to others
daa_learning_status- Show learning progress of autonomous agents
daa_cognitive_pattern- Read or change an agent's cognitive pattern
daa_meta_learning- Transfer learned behaviour between domains
daa_performance_metrics- Report metrics for the autonomous-agent layer