Runtime and API reference
Fleet has two programmatic boundaries: the dashboard's authenticated workflow API and the local MCP server. Fabric remains a coordination and audit rail, not the workflow execution API.
Workflow run states
| pending | Queued or requeued for a worker. |
| awaiting_code | Parked while a code step waits for its PR. |
| awaiting_approval | Parked while a human reviews the gate evidence. |
| done | Completed successfully. |
| failed | Execution failed or exhausted its recovery path. |
| rejected | A human rejection ended the run. |
Workflow API boundary
The browser uses authenticated definition, run, approval, artifact, and history routes. The local worker uses instance-token routes for pending runs, leases, trigger definitions, completion reports, and code results. These are deployment internals rather than a public unauthenticated REST API.
User-authenticated
Definitions, manual enqueue, run history, approval/rejection, artifacts, and intervention acknowledgement.
Instance-authenticated
Pending/active runs, run claims, trigger enqueueing, completion, notifications, and PR code results.
MCP transport
fleet mcp serve
Transport: JSON-RPC 2.0 over stdin/stdout
Discovery: tools/list
Invocation: tools/callThe client starts Fleet as a child process. There is no MCP HTTP port. Use discovery rather than a hard-coded tool count because the registered surface evolves with the binary.
Agent environment
| FLEET_AGENT_NAME | Named agent instance. |
| FLEET_AGENT_ROLE | Agent type / role. |
| FLEET_AGENT_DEPARTMENT | Department metadata. |
| FLEET_AGENT_REPORTS_TO | Manager metadata. |
| FLEET_PROJECT | Current owner/repo project scope. |
| FLEET_REPOS | Comma-separated org repositories for an org agent. |
| FLEET_OWNER | Human owner used for escalation. |
| FLEET_SESSION | Fleet-managed session name. |
| FLEET_WORKDIR | Repository or isolated worktree used by the agent. |
| FLEET_TASK_TITLE | Authoritative task instruction supplied by assignment or workflow dispatch. |
| FLEET_TRIGGER_EVENT | Launch context such as task_assigned. |
| FLEET_TRIGGER_PAYLOAD | Structured context for the assigned task. |
| FLEET_PUBLISH_CMD | Pre-scoped command for publishing a coordination event. |
| FLEET_PRODUCT_NAME | Configured product name. |
| FLEET_PRODUCT_DESCRIPTION | Configured product description. |
The removed pipeline-stage environment variables are not part of current workflow dispatch. Workflow instructions arrive through the assigned task and persisted run context.
Host controls
FLEET_DB_PATH overrides the local database path.
FLEET_GENFLOW_MAX_RUNS caps agent-running steps in one run.
FLEET_GENFLOW_MAX_FANOUT caps items produced by a fan-out step.
FLEET_GENFLOW_ALLOWED_RUNNERS extends the trusted runner allowlist.
FLEET_GENFLOW_EVENT_TRIGGERS=0 disables label-trigger polling for migration or debugging; schedules remain independent.
Data paths
| ~/.fleet/fleet.db | Local SQLite operational database. |
| ~/.fleet/logs/ | Watcher and agent logs. |
| ~/.fleet/orgs/<name>/org.yaml | Named organization configuration. |
| ~/.fleet/watcher*.pid | Repo watcher PID files. |
| ~/.fleet/org-*-watcher.pid | Org watcher PID files. |
| ~/.fleet/brain.sock | Brain daemon socket. |
| .fleet/config.yaml | Repository-local worker, agents, types, corpus, and connector config. |
| .fleet/prompts/ | Repository prompt overlays. |
| ~/.claude/skills/fleet/ | Installed Fleet skill destination. |