Skip to main content
Use Codex, Claude Code, Pi, or a mix. Flowfield includes their integrations; native accounts, providers and advanced settings stay with each harness.

Setup

  1. Install and sign in to a supported harness on the machine running Flowfield.
  2. Choose the Coordinator’s harness/model beside its input. Choose worker defaults in Workspace settings → Workers; tasks can override them.
  3. Configure and validate project integration before running workers.
Settings → Harnesses detects installations. Override executable or configuration-directory paths only when detection needs help. Verify installation checks version and native sign-in for Codex/Claude Code; Pi verification checks version, while provider access remains in Pi. It sends no model prompt. Successful use establishes model access. Verification results persist across page reloads while the service runs. Verify again after installing/upgrading a harness or changing its account or paths. Unavailable harnesses show their setup remedy and leave ready harnesses usable. Models load when you open a selector; saved choices preload on open and reconnect. Discovery opens a disposable native session without a model prompt, so startup hooks can run. Harness changes clear dependent model, effort, access and Fast choices. Native capabilities are documented on each agent page.

Flowfield tools

Built-in Coordinator and workers receive MCP connections automatically, scoped to their workspace or task attempt. Coordinator repository operations name a member project; workers retain their assigned checkout and selected context. Browser, CLI and tools share the same revision, saved-answer and exact-approval rules. Native instructions and extensions can still apply. Reconcile relevant conventions through workspace documents; selecting an access mode or allowing a tool never approves worker code delivery.

Standalone conversations

To coordinate from your own harness, connect it to the running service. Use codex, claude-code or pi:
Connect saves a user-scoped native MCP connection while preserving other servers. Status checks configuration, service reachability and tool discovery. Start a fresh harness conversation to load it; native project overrides and organization policy still apply. Optionally install the coordinator skill in an adopted project:
This adds a skill and AGENTS.md reference while preserving existing instructions. Review and commit the files. Ask the fresh conversation to read the guidance, find the intended workspace and inspect its projects and board. Managed task workers use their own assignment instructions. For a different service port or connection name, use matching options:
Disconnect removes that connection; restart the harness to unload its tools. Work records remain saved. Built-in roles need none of these standalone commands.

Recover setup and discovery

Follow a failed turn’s recovery message, check native sign-in/provider access on the service host and resend deliberately. Flowfield never automatically replays an interrupted prompt. Harness-specific remedies live on the Codex, Claude Code and Pi pages. A lost model-discovery owner can leave a hold across restart. Inspect and stop remaining native discovery on the host, then use the confirmation in Harnesses settings or the CLI with the exact discovery ID. Confirmation clears the hold without starting a process; refresh models afterward. For an interrupted Coordinator session, see Stop and continue later. Worker recovery actions appear in tasks and Inbox.