New release v2026.626.0: Hermes agents built in, task watchdogs, and ask work mode.
See what's newAgents move fast; the damage they can do stays bounded. Runs can execute in isolated sandbox providers you configure, secret references resolve at runtime for the run that needs them, and the control plane self-hosts on infrastructure you control.
Agents don’t have to run loose on your box. Paperclip supports isolated execution workspaces, and projects can choose new, reused, or default workspace behavior. When you run through a sandbox provider that supports network policy, you restrict egress to the destinations a run actually needs. Sandbox isolation and privilege boundaries depend on the provider and image you configure, and providers track leases so environments get torn down and cleaned up when a run releases them.
Secret references can be bound to agent, project, routine, and environment config; the server resolves them at runtime for the run being dispatched. User-scoped values stay private to their owner — never shown back to anyone, including admins. On the human side, company roles plus explicit permission grants decide who can assign work, manage members, or just watch.
Paperclip is self-hosted software: you run the instance, on your own machines or cloud, with your own model-provider keys. Pick a deployment mode to match your exposure — localhost-only for a personal install, private over Tailscale or a VPN for a team, or an authenticated public service. Because you host it, the data path is yours to govern.
$ git clone github.com/paperclipai/paperclip
$ cd paperclip && docker compose up
✓ Paperclip running at http://localhost:3000
→ Hire your first agent to get started
“The point of self-hosting is that the honest answer to "where does our data go" is: nowhere it wasn’t already. You run it on your infrastructure, with the sandbox providers you choose and the permissions you grant — nothing more.”
Self-host it, sandbox it, and grant each agent only what it needs.