paperclip-mcp

An MCP bridge to Paperclip. Read your companies and issues, and act on what awaits a human, from any remote MCP client, with a policy on every tool.

Source on GitHub

A morning round, handing work to an agent, what gets refused, and the path each request takes. Fictional data.

Why

The official @paperclipai/mcp-server speaks stdio and exposes everything, approvals included, plus raw API access. A board token has every right on every company: this bridge exposes a closed set of tools instead, and is the only thing that restricts that token.

One mode per tool

allowThe tool runs.

askThe server asks you to confirm through MCP elicitation. A client that cannot ask is refused.

denyThe tool is not exposed. Default for anything not listed.

Decision tools only accept ask or deny; the server refuses to start otherwise.

Tools

`companies`	Companies, issue prefix, id
`attention`	What awaits a human, across companies
`issues` · `issue` · `document`	Issues, detail, comments, documents, pending cards
`agents`	Agents of a company
`approvals` · `action_requests`	Pending approvals and tool actions
`issue_comment`	Comment on an issue
`issue_create`	Create an unassigned issue, optionally in a project
`issue_update`ask	Assign an issue to an agent or to yourself, change its status, with an optional comment. Assignment and safe statuses (backlog, todo, done, cancelled by default) run directly; other changes ask
`projects` · `project_create`	List projects, create one
`runs` · `run`	Runs in progress, an agent's last runs, a run's error and last messages
`issue_history`	What happened on an issue
`spend`	Tokens and runs per agent over the last days
`activity` · `routines`	Latest events of a company, its routines
`agent_pause` · `agent_wake` · `run_stop`	Pause an agent, wake it, stop a run
`approval_comment`	Comment on an approval without deciding it
`agent_resume` · `agent_unblock` · `routine_run`ask	Resume an agent, clear its error, run a routine now
`approval_decision`ask	Approve or reject an approval
`interaction_respond`ask	Answer an agent's question card
`plan_confirm`ask	Accept or reject a plan
`action_request_decision`ask	Authorize or decline a suspended tool action

Making an agent act

In Paperclip, a comment only wakes the assigned agent, and an issue assigned to a user wakes nobody. To hand work to an agent, issue_update assigns it, sets the status to todo and adds a comment saying what to do, in a single request.

Run

docker build -t paperclip-mcp .
docker run -e PAPERCLIP_API_KEY=... \
  -e CF_ACCESS_TEAM_DOMAIN=... -e CF_ACCESS_AUD=... \
  -v ./policy.yaml:/app/policy.yaml:ro -p 9200:9200 paperclip-mcp

Every request must carry a valid Cloudflare Access JWT: a shared Docker network is not enough to drive Paperclip. The MCP endpoint is /mcp.