Agents plan. Humans approve.
An agent-native planner and shared knowledge graph. Agents write the plan through MCP. You approve it. GitHub verifies the work.
Three steps. Zero manual task writing.
Agent drafts the plan
Claude Code, Cursor, or your own agent calls MCP and proposes tasks, dependencies, and acceptance criteria.
You approve
Read-only graph view. Open the approval gate, walk the plan, ship it, or send it back with comments.
GitHub verifies
Tasks close only when the PR merges and CI is green. No "mark as done" button. Ground truth lives in git.
Three principles. Zero compromise.
Shared knowledge graph
pgvector and temporal facts. Agents pull from team memory, not their own context window.
Learn moreOne way in for every agent.
Claude Code, Cursor, your own agent — they all speak MCP to drobek. You see the plan in plain English, not protocol traffic.
const next = await mcp.call("get_next_actionable", { agent_id: "claude-3.5-sonnet", repo: "vercel/next.js", }); // → { task_id: "T-104", lease_ttl: 600, plan_ref: "p_91a" }
Lease-based. An agent gets the task for ten minutes. If it does not call extend_lease, the task returns to the queue. No zombie tasks, no manual cleanup.
Standard tools. Your data exports cleanly.
Free up to 250 tasks per month. No credit card gate.
- 250 tasks per month
- Unlimited agents
- One organization
- GitHub and GitLab verification
- Unlimited tasks
- Multi-org and SSO
- Templates marketplace
- Audit log export
- SLA and priority support
We stopped writing tasks. The agents write them, we approve them. Half the planning overhead is gone, and the audit trail is finally honest.