Planner — provable "done" for AI agents

Generating code got cheap. Verifying it didn't. Every time an AI agent says "done", someone re-checks it by hand — that's verification debt. Planner is a hosted MCP server that makes "done" provable instead of trusted.

How it works

1. Decompose — break a goal into a tree of subtasks, one level at a time.
2. Acceptance criteria — every node carries behavioral, objectively-checkable criteria (no "looks good to me").
3. Evidence-gated done — a node can't close until real evidence is attached for each criterion, and an independent LLM judge checks that evidence and accepts or rejects it.

Why it's different

Most task MCP servers structure work and then mark it done on trust. Planner closes the loop: the agent proves each criterion; a judge verifies the proof. Task trackers tell you what to do — Planner tells you whether it was actually done.

Features

• Iterative decomposition with behavioral acceptance criteria
• Evidence model: attach artifacts per criterion + independent judge verdict
• Goal gate-keeper interview on creation (why / where / how we'll verify)
• Projects, dependencies, priorities, tree & summary views
• Hosted & remote: streamable-http with OAuth 2.1 + Dynamic Client Registration — nothing to install

Connect

Remote MCP endpoint (OAuth 2.1, DCR supported):

{
  "mcpServers": {
    "planner": {
      "type": "http",
      "url": "https://planner.monopoly-gold.com/_mcp"
    }
  }
}

Learn more: https://planner.monopoly-gold.com