§ Changelogmcp-0-2-0

MCP 0.2.0: a human approval gate for any agent host (mcp-0-2-0)

`agentdraft-mcp` 0.2.0 adds request_approval, get_approval, and wait_for_approval, plus an interactive config generator on the MCP page and the approval guides.


Upgrade

pip install -U agentdraft
npm install agentdraft@latest

See each package's ## Upgrade section on its registry page for the package-specific command.

What changed

  • Added

    agentdraft-mcp 0.2.0 exposes nine tools: the six scheduling tools plus request_approval, get_approval, and wait_for_approval. The wait is clamped to 120 seconds, and a timeout returns status pending rather than a false approval.

  • Added

    An interactive gate console on /for-agents/mcp and every human-approval field note writes the exact MCP config for Claude Desktop, Claude Code, Cursor, or Cline, with uvx, pipx, pip, or python -m. The optional API key stays in browser memory and is never stored or sent.

  • Added

    agentdraft-benchmark is now a standalone MIT package with a public mirror at ryabinski-labs/agentdraft-benchmark, a committed baseline report, and --keys-file / AGENTDRAFT_AGENT_KEYS support so anyone can re-run it against production without dev-only signup.

  • Fixed

    The duplicate H1 shared by the two webhook replay-protection posts is resolved: each now owns a distinct title and description, pinned in both the publish-time renderer and the prebuild sync so a content run cannot revert them.

  • Fixed

    The approval workflow guide carries a technical title, H1, and meta description that name async gates, JSON evidence, and MCP — the mechanisms the article actually covers.