Skip to main content
MCP Documentation

Orchestrator Service Guide

Plan, execute, and verify multi-step work with approvals, autonomy budgets, and full audit trails managed locally.

Execution flow

Goals are parsed into a directed acyclic graph (DAG). Each node references a tool contract, budget, and approval policy before the orchestrator enqueues the step.

  • `/tasks` creates a run request with context, policy mode, and budget caps.
  • `/tasks/{id}/run` streams status updates so interfaces can visualize plan progress.
  • Retries are automatic when tools return retriable errors; fatal faults pause for review.

Approvals & safety

High-risk steps transition into a `pending` state until an operator approves through the CLI or upcoming GUI.

  • Policy modes (`constrained`, `review_required`, `autonomous`) determine which steps block.
  • Autonomy budgets stop runaway tasks; thresholds trigger alerts in the episodic log.
  • All actions are timestamped and attributable for compliance evidence.

Verification hooks

Before producing final output, orchestrator workflows can require evaluation checks such as `mcp-eval` tests or rubric scoring.

  • Failed checks automatically branch into remediation nodes or request human review.
  • Success criteria and artifacts are stored with the task metadata.
  • Outputs can be exported with provenance for downstream auditing.