Local MCP beta

FlowOS

FlowOS is a by-the-book local MCP server that gives an agent durable records, artifacts, capabilities, actions, and run evidence.

Download local package Bootstrap Privacy Support

Beta access: FlowOS Local MCP is being tested now. For access notes, feedback, or beta coordination, email [email protected].

What it does

FlowOS runs next to the agent as a local stdio MCP server. It does not run a model, host your data, or require an Offband account. Goose, Codex, or another MCP-capable client starts the server and calls its tools during normal work.

The server stores state in SQLite under ~/.flowos. Each workspace has its own database, so a local model or agent can keep project memory separate without depending on hosted infrastructure.

Core objects

  • Records keep decisions, evidence, issues, notes, standards, and handoffs.
  • Artifacts store markdown, JSON, command output, configs, and other useful work objects.
  • Capabilities describe reusable procedures an agent can validate and improve.
  • Actions define manual, command, or script steps inside a capability.
  • Runs preserve execution evidence from using a capability or action.

Bootstrap

Download the local package for your machine, unpack it, initialize a workspace, and point your MCP client at the binary.

# Linux x86_64 curl -L -o flowos-local-go.tar.gz https://flowos-api.offband.dev/downloads/flowos-local-go-linux-amd64.tar.gz # Linux arm64: replace linux-amd64 with linux-arm64 # macOS: use darwin-arm64 or darwin-amd64 tar -xzf flowos-local-go.tar.gz cd flowos-local-go ./flowos init scripts/local-flowos-userland-smoke.sh ./flowos mcp

Agent bootstrap prompt

Give this to a zero-context agent after adding the MCP command:

You have a local FlowOS MCP extension. At the start of meaningful work, call flowos.agent_bootstrap with the task. Use get_relevant_context to recover prior context. Use find_capabilities before drafting new procedures. Use list_actions before run_action. Write durable findings, decisions, issues, run evidence, and handoffs with flowos.capture_learning or submit_record.

Deferred

The public FlowOS skill registry and shared skill marketplace are not part of the current beta. The product focus is the local MCP server, its durable object model, and the agent loop for collecting, running, validating, and refining capabilities.