ruv-swarm MCP Server: Setup for Claude Code, Cursor and Other Apps

• Starts a swarm of agents with a layout you choose

  Creates a new swarm on your machine. You pick the layout (mesh, hierarchical, ring or star), a cap of 1 to 100 agents and how work is shared. The swarm is saved to a local database.

  Try: “Start a mesh swarm with up to 8 agents.”

  Write · swarm_init

• Adds an agent to the swarm

  Adds one agent of a type you choose (researcher, coder, analyst, optimizer or coordinator), with an optional name and list of skills. The agent is saved to the local database.

  Try: “Add a researcher agent called "Docs reader".”

  Write · agent_spawn

• Hands a task to the swarm

  Records a task, assigns it to agents in the active swarm and returns a task id. You choose a parallel, sequential or adaptive plan and a priority, and the task is saved to the local database.

  Try: “Run "audit the auth module" across the swarm, in parallel.”

  Write · task_orchestrate

• Shows the current swarm and its agents

  Reports the active swarm and its agents. Ask for the verbose form to see each agent.

  Try: “What is the status of my swarm?”

  Read · swarm_status

• Lists agents, filtered by state

  Lists the agents in the swarm. You can show all of them or only active, idle or busy ones.

  Try: “Which agents are idle right now?”

  Read · agent_list

• Shows performance numbers for agents

  Returns metrics for all agents or one agent id, including stored neural network details.

  Try: “Show metrics for the coder agent.”

  Read · agent_metrics

• Checks how running tasks are going

  Reports progress for running tasks, or for one task id, with an optional detailed view.

  Try: “How far along is task task-123?”

  Read · task_status

• Fetches what a finished task produced

  Returns the stored result of a completed task by id, as a summary, in detail or raw.

  Try: “Show me the full results for task task-123.”

  Read · task_results

• Watches swarm activity for a few seconds

  Samples swarm activity at an interval you set for a set time, 10 seconds by default. It also logs a monitoring event to the local database.

  Try: “Monitor the swarm for 30 seconds.”

  Write · swarm_monitor

• Times the swarm engine on your machine

  Runs speed tests on the WebAssembly engine, swarm creation, agents and tasks for up to 100 iterations. It uses your CPU while it runs.

  Try: “Run all benchmarks for 20 iterations.”

  Runs code · benchmark_run

• Checks what your machine supports

  Reports which runtime features are available, such as WebAssembly, SIMD, memory and platform details.

  Try: “Does this machine support SIMD?”

  Read · features_detect

• Shows how much memory the swarm uses

  Reports memory use for the WebAssembly engine and JavaScript, in summary, in detail or per agent. It also stores those readings in the local database.

  Try: “Show memory use by agent.”

  Write · memory_usage

• Shows neural network state for agents

  Reports the status and performance of the small neural networks attached to agents, for all agents or one.

  Try: “What is the neural status of the researcher?”

  Read · neural_status

• Trains the agents' neural networks on sample tasks

  Runs 1 to 100 training rounds on sample tasks for one agent or all of them. It saves the updated networks and the training numbers to the local database.

  Try: “Train the coder agent for 20 iterations.”

  Runs code · neural_train

• Describes the thinking styles agents use

  Returns information about cognitive patterns such as convergent, divergent, lateral, systems, critical and abstract thinking.

  Try: “Describe the lateral thinking pattern.”

  Read · neural_patterns

• Starts the autonomous agent (DAA) service

  Turns on the DAA service, with options for learning, coordination between agents and whether state is kept in memory or on disk.

  Try: “Initialize the DAA service with learning enabled.”

  Write · daa_init

• Creates an autonomous agent

  Creates a DAA agent with an ID, a list of capabilities, a thinking style and a learning rate.

  Try: “Create a DAA agent called reviewer with a critical thinking style.”

  Write · daa_agent_create

• Adjusts an agent from feedback

  Sends feedback, a score from 0 to 1 and suggestions to an agent so it changes how it works.

  Try: “Tell agent reviewer its last result scored 0.6 and it should be briefer.”

  Write · daa_agent_adapt

• Defines a multi-step workflow

  Creates a workflow from steps and their dependencies, run in parallel, in order or adaptively.

  Try: “Create a workflow with three steps that run in sequence.”

  Write · daa_workflow_create

• Runs a workflow

  Runs a saved workflow with the agents you name, optionally in parallel.

  Try: “Execute the workflow using agents reviewer and writer.”

  Write · daa_workflow_execute

• Passes knowledge between agents

  Copies knowledge in a named domain from one agent to one or more others.

  Try: “Share the testing knowledge from reviewer to writer.”

  Write · daa_knowledge_share

• Shows how agents are learning

  Reports learning progress for one agent or all of them, with an option for more detail.

  Try: “Show detailed learning status for all DAA agents.”

  Read · daa_learning_status

• Analyzes or changes an agent's thinking style

  Either analyzes an agent's cognitive pattern or switches it to convergent, divergent, lateral, systems, critical or adaptive.

  Try: “Switch agent reviewer to the lateral pattern.”

  Write · daa_cognitive_pattern

• Moves learning between domains

  Transfers what agents have learned from one domain to another, in adaptive, direct or gradual mode.

  Try: “Transfer learning from testing to documentation, gradually.”

  Write · daa_meta_learning

• Shows DAA performance numbers

  Returns metrics for the DAA service by category and time range, such as 1h, 24h or 7d.

  Try: “Show all DAA metrics for the last 24h.”

  Read · daa_performance_metrics

Watch out for

• It writes swarm state to a local database

  Creating swarms, agents and tasks, training networks and some monitoring calls save to a SQLite file (data/ruv-swarm.db) inside the package folder. Delete that file if you want a clean slate. github.com

• Three more tools work than the server lists

  The server lists 25 tools: 15 core tools and 10 DAA (autonomous agent) tools. It also accepts calls to three others by name: pool_health, pool_stats and persistence_stats, which report on the database connection pool. Your assistant will not see them unless it already knows the names. github.com

• Benchmarks and training use your CPU

  Benchmark and training calls run for up to 100 rounds on your machine. Ask for small numbers first. github.com

• The README's environment variables are not read

  The README sets RUV_SWARM_LOG_LEVEL and RUV_SWARM_ENABLE_SIMD in one example and SWARM_CONFIG and MAX_AGENTS in another. The 1.0.20 server code does not read any of them. Its logger reads LOG_LEVEL, LOG_TO_FILE, LOG_FORMAT and LOG_DIR instead. github.com

Security advisories