IMPLICATOR
Implicator PRO Briefing / 29 Sep 2026
Paperclip organizes AI agents into a working team. Our trial found useful controls alongside defaults that let agents hire colleagues, skip permission prompts and inherit the host’s environment.
Paperclip is free, self-hosted software for managing AI agents as a team. It contains no AI model. The owner brings agents such as Claude Code or Codex, gives them roles and tasks, and can put a CEO agent above other workers. Budgets and approvals govern parts of their activity, while the human owner sits above the organization as “the board.”
Implicator tested the controls on September 28, 2026, using stand-in agents that exercised Paperclip without calling an AI model. The owner’s powers held: an agent could not raise its own budget, read company secrets or approve requests, and a human pause stopped our running job. A run that Paperclip classified as Claude subscription use reported $5.00 and counted $0.00 against a $1.00 cap. Claude Code agents started with permission prompts switched off. A new agent could hire another until the owner changed a setting.
These settings matter when an agent works through a company account, runs on a developer’s laptop or shares a server that holds keys. This PRO briefing reports what the current release did in our test and gives operators an 11-step pre-flight checklist.
The short version
- Paperclip is free, self-hosted software that runs AI agents like a company, with a CEO agent, budgets, approvals and a human “board.”
- In our September 28 test, a run Paperclip classed as subscription use reported $5.00 and counted $0.00 against a $1.00 cap.
- Claude Code agents launched with permission prompts skipped, and a new agent hired another agent until board approval was switched on.
- Agents could not raise budgets, read company secrets or approve requests, and pause stopped a running job within a second.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
What Paperclip does
The MIT-licensed Paperclip project gives an agent team an org chart. Tasks carry assignments and comments. Scheduled check-ins, called “heartbeats,” wake agents to look for work, while routines handle recurring jobs. Cost limits, approvals and an activity log let the owner see what ran and intervene. An agent’s ability to complete an assignment comes from its own model, tools and access. Paperclip supplies the management layer around them.
The repository was created on March 2, 2026. On September 28, it had 92,720 stars and 5,989 open issues. Stars indicate interest; they do not establish installations or successful deployments. The issue count is a backlog at that date, not a count of security defects. Paperclip can be self-hosted without an account, though operating it still requires a machine, an agent runtime and the keys that runtime needs for its assigned jobs.
The pseudonymous co-founder Dotta describes two problems facing agent teams: “Quality and governance permissions. Those are the two biggest things.” He wants owners to decide which agents may search, forward or reply to email and to see a record of their tool calls.
In a video published September 24, IT educator Chuck Keith gave a newly hired agent access to a security service through Paperclip’s secrets screen. It soon assigned another agent forced password resets Keith had never requested, and he reached for Paperclip’s pause control.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
