Preview access

Code review that actually runs the code. Compute that scales quality, not quantity. All on the cloud, built for agents.

I'll run Orpheus to spin up parallel agents on isolated VMs for this task.
Bash orpheus exec spec.md --program orchestrator
→ orchestrator on vm-a3f2
  → claude-reviewer on vm-b1c4 (fork, 1.2s)
  → codex-executor on vm-d8e1 (fork, 0.9s)
→ agents working...
→ PR #142 opened gentle-nocturne
PR is up. The orchestrator picked the Claude solution over Codex after comparing test results.
Fix OAuth session handling and add token refresh #142
Open orpheus[bot] wants to merge into main from orpheus/gentle-nocturne
O orpheus[bot] 2 minutes ago
E2E run caught a redirect loop on the OAuth callback when the session cookie is expired. Fixed by checking cookie validity before redirect and falling back to a fresh auth flow. Added regression test in test_oauth_e2e.py.
+3
All checks passed

Agents lie.

To you, and to themselves. They'll write hollow tests, declare victory, and never look back. The problem is they rarely have the tools to verify their own work. And even when they do, they don't bother.

Our solution.

Infrastructure

You define a program — orchestrator, sub-agents, roles — and Orpheus provisions VMs, allocates agents, manages forks. Bring whatever agents you want. We'll manage the fleet.

Verification

Orpheus invests compute in validating code, not just producing it. Agents are required to run code end to end, execute the happy path, and review outputs before anything gets merged. We use defence in depth to keep agents honest.

Flexible by design.

Work locally with Claude, Codex, Gemini, or anything else you prefer. Point them at the Orpheus CLI and they become commanders of the agent fleet — launching test-time compute on fully provisioned machines and delivering reviewed, verified pull requests. Add it to any workflow.

Get started

Quickstart