Orcho makes AI work accountable

Read first

Run one task. Watch Orcho plan, implement, reject false-ready work, repair it, and prove what is ready to deliver. AI agents remain the workers. Orcho is the production harness and control plane around them: profiles define the work shape, gates decide whether it can continue, and evidence keeps readiness inspectable.

intentprofilepipelinegatesevidence

Start quickstart5-minute story

live CLI firstmono-run firstdeep reference later

orcho run —output liveoperator view

Orcho Run  20260629_1427  profile=feature
Task  Add validation to the login endpoint

Pipeline
⟳² (▶ plan [Claude] → · validate_plan [Codex]) → · implement [Claude]
  → ⟳² (· review_changes [Codex] → · repair_changes [Claude])
  → · final_acceptance [Codex]

[PLAN] accepted implementation contract
[IMPLEMENT] edited api/auth.py and tests
[REVIEW_CHANGES] verdict=REJECTED
blocker  missing negative-path test
next     correction follow-up

Pipeline glyphs

⟳² = loop of up to 2 rounds · ▶ = active phase · · = waiting · [Claude]/[Codex] = worker runtime

Primary perception

The terminal stream shows phase progress, agent output, findings, and next action while the run is alive.

Afterwards

orcho status and orcho evidence help inspect the persisted record.

Artifacts

output.log, events.jsonl, plan.md, review.json, diff.patch

01Run one taskInstall Orcho, create a workspace, run a mono task, and inspect the result.02Watch the live runThe CLI stream is the first operator UX.03Improve qualityUse profiles, gates, receipts, and evidence to catch false-ready work.04Measure progressEvaluate quality, cost, runtime mix, proof, and operational risk over a run window.

orcho run · a real run, plan rejected then re-planned, shippedanimated

A real run, in full: the plan is built as a delivery contract — acceptance criteria, owned files, risks, six tasks — rejected on review with a concrete finding, re-planned, approved, then executed as a six-subtask DAG where every subtask states its goal and done-criteria and closes with a per-criterion attestation, through review and final acceptance to a delivery commit. It ends on the accountability rollup: ~$18 API-equivalent, 31 minutes hands-off, 6/6 tasks, one finding raised and resolved, every gate backed by a receipt. Pause or scrub the player; the full-pace walk-through is on Watch the run.

Semantic plane

A profile tells Orcho what kind of work this is

task intentprofilephasesgatesevidence

Do not read Orcho as “run many agents.” Read it as “choose the operating shape for a task.” The same request can be a light mono-run, a gated implementation run, a review-only run, or a participant-set workflow. The profile makes that choice explicit before the lifecycle starts.

Understand profile semantics

Choose your level

Read from light entry to expert control

01Start simpleUnderstand Orcho, run one mono task, watch live output, and read the result.02Improve qualityChoose workflow depth, catch false-ready work, and require durable receipts.03Measure and governTrack cost, review load, evidence strength, recovery, and adoption over time.04Build and extendShape profiles, gates, workspaces, prompt composition, runtime adapters, and low-level contracts.

Choose your sub-domain

Orcho has separate operating planes

Run controlLifecycle, phase gates, follow-ups, resume, and delivery decisions.Profiles and gatesWorkflow depth, readiness policy, false-ready delivery, and verification receipts.Evidence and trustStatus, evidence bundles, receipts, artifacts, events, diffs, and final acceptance.Measurement and governanceQuality, cost, runtime economics, proof, and operational risk across a run window.Workspaces and participantsRun state, generated worktrees, one project first — then several repositories delivering one change together.Cross-project modeOne feature intent, project aliases, per-project handoffs, contract check, and one system gate.Worker runtimesClaude, GLM-compatible wrappers, Codex, Gemini, phase routing, runtime readiness, and adapter boundaries.Prompt engineRole/task/format composition, protected contracts, project overrides, and utility-lab primitives.Skill registryPortable capability packages, trust policy, subtask routing, and skill-use evidence.MCP and referenceTyped run control for clients plus CLI, artifact, and event contracts.System designEngine model, recorded run state, gates and verification, isolation, and the MCP control plane.