Approving
Agent workflows with human collaboration — a new multi-agent paradigm.
Approving turns coding agents into trusted workflows: visual orchestration, sandboxed execution, and human Approve gates at critical steps.
- Cursor
- Claude Code
- CodeBuddy
- Trae
Human–agent collaboration, made simple
Review is the bottleneck in multi-agent work. Visual, interactive approvals make gates what-you-see-is-what-you-get.
human_gate · waiting_human
Sign-in · UI DEMO
Review an interactive demo — annotate issues, then reject or approve.
Welcome back
Continue with your work account
Renew session
Stay signed in — confirm securely
Open the approval window…
Project management, redefined
Every project gets a PM. Connect your usual IM and manage work from your phone.
Next: richer org roles — director, GM, and beyond. OPC, for real.
ApprovingPM
Workflow PM · chat is orchestration
- Harness: eval suite + regression bar proposed
- Engineering: coverage / lint / security in CI proposed
- Gates: human Approve on critical paths proposed
- Harness eval scaffolding queued
- CI quality gates queued
- Wire human Approve queued
Parallel scheduling, faster iteration
Agents run in sandboxes, coordinated by the PM — iteration gets another gear.
ApprovingPM
IM · start workflows by chat
IM confirmed — running in parallel…
Capabilities
- Visual orchestration Canvas FSM with live states; parallel requirement lanes and a Gates Inbox.
- Real Docker sandboxes Execute in containers via ACP and sandbox-gateway; the Web UI uses the API.
- Multi-agent backends Cursor, Claude Code, CodeBuddy, Trae — pick acpBackend per agent.
- Artifact contract + MCP write_artifact / set_* / node_complete — isolated by run token.
- Git credentials in sandbox GITHUB_* / GITLAB_* / SSH; GitLab auto-MR and GitHub gh supported.
- Single-repo self-host sandbox-gateway and sandbox image sources live in this repo — one clone to self-host.