Drive a team of AI agents from Claude, without handing it the decisions

The need

Paperclip's board token can do everything: approve a plan, hire an agent, call any API endpoint. The official MCP server exposes all of it, and only locally. I wanted to follow and nudge the work from Claude, wherever I am, without giving it that power.

What it does

• The daily picture: what waits for a human across every company, whether an approval, an agent's question, a review or a failure.
• A task in one sentence: created in the right project, assigned, moved to todo, with the brief as a comment. The agent wakes up and starts.
• Project follow-up: issues by status, conversations, the plan written by the agent.
• Nudge an agent with a comment, or clean up finished issues.

How it is protected

• Every request goes through Cloudflare Access, and the bridge checks the signed token itself (RS256, audience, issuer, expiry). Being on the same Docker network is not enough.
• The Paperclip token stays inside the bridge. Claude only sees a closed list of tools, each with a mode: run, ask for confirmation, or not exposed.
• Decisions (approving a hire, answering an agent's question, accepting a plan) require explicit confirmation or are not exposed. The server refuses to start if they are opened up.
• For issues, the rule goes down to the change: assigning, or moving to backlog, todo, done or cancelled runs directly. Other statuses ask for confirmation.

Limits

claude.ai and the Claude app do not show MCP confirmation prompts yet. Actions that need one are refused there, and nothing is written. That is by design: a refusal beats an implicit yes. Hence the fine-grained rule that lets safe changes through.