FlowJoe is designed for both human and agentic collaboration. Through the Agent Bridge, FlowJoe hosts a local Model Context Protocol (MCP) server, exposing your outlines, workspaces, and media galleries to external terminal-driven agents.

​
How It Works

The Agent Bridge acts as a secure local loopback server. When active, it translates FlowJoe’s complex visual structure and files into a standard format that external LLM agents can read and write:
  • Read Snippets: External agents can read your active tree structure, notes, and task states.
  • Propose Changes: Agents can use the standard flow-mutation verb set to write new nodes, stage images, or edit checklists.
  • User Consent: The agent cannot change your project directly. Every command is sent as a proposal card to your FlowJoe review inbox.

​
🚀 Setting Up the Bridge

  1. Open Settings → Agent Bridge (MCP).
  2. Toggle the Bridge Server to On. The server starts locally on your machine.
  3. Bearer Token Security: FlowJoe generates a secure bearer token to authorize local connections. You can rotate this token at any time.
  4. Direct Connect Integration: Click the Direct Connect button to automatically write connection files to your local tool config directories (e.g. ~/.config/claude/mcp.json), making Joe immediately available to tools like Claude Code or custom Python runtimes.

​
🛡️ The Automation Seatbelt

When an external terminal agent attempts to apply updates, FlowJoe passes the changes through the Automation Seatbelt:
  • Non-Destructive Edits: Adding cards, appending notes, or creating tags can be applied automatically if authorized.
  • Destructive Intercepts: If a script attempts to delete a node, delete task logs, or remove content it does not own, the seatbelt halts the run. The proposal is redirected to your manual review inbox, ensuring you stay in absolute control.