BLOG
Agents plan. Humans approve. Introducing drobek.
The problem: agents are write-blocked from your planner
Most engineering teams adopting AI agents end up with a split-brain workflow: the agent thinks in tasks and pull requests, but the planner of record (Linear, Jira, Notion) only accepts writes from humans. You get screenshots pasted into Slack, hand-copied task titles, and an audit trail that ends at the chat window.
How drobek inverts the flow
drobek is MCP-first. Your agent uses propose_plan, add_task, update_task_status directly. The planner accepts the writes; humans review. A plan lands as proposed and you flip it to active with one click. No forms to retype, no screenshots.
What humans still do
Three things, deliberately:
- Approve proposed plans. A plan does not become "active" without you.
- Resolve `needs_review` tasks. If verification can't prove the work, it lands here, not at
done. - Suspend, redirect, or comment. You stay in the loop without being a typist.
Why GitHub stays the ground truth
A task does not transition to done because the agent says so — it transitions because the PR merged, the expected files changed, and CI is green. The verification engine is the gate. Anything ambiguous goes to needs_review and you decide.
Try it: 3-minute Claude Code setup
Sign up, create a workspace, then run claude mcp add --transport http drobek https://mcp.drobek.app/mcp. Claude Code opens a browser tab, you sign in (OAuth 2.1), and you're connected — no tokens to copy. Ask it to "use the drobek MCP server to propose a plan to add Apple Sign In as an OAuth provider." The plan lands in /plans for your approval.