Get started

What is Paperclip?

Paperclip is the governance layer for AI agents. Roles, reporting lines, budgets, approvals, audit trails — the actual data model of an organization. 5dive is the runtime layer underneath. Processes, identities, channels, isolation tiers — the substrate that any agent, governed or not, has to live on. Different floors of the same building.

For most teams, the runtime floor is enough. 5dive alone handles agent coordination, messaging, scheduling, isolation — the day-to-day of running a small AI team — without any of the governance ceremony Paperclip adds on top. The rest of this post explains what Paperclip is and when it earns its keep, with the honest answer that for the majority of workloads it’s overkill.

A quick disclosure up front: Paperclip is not our project. We are writing about it because it solves a problem worth knowing about, and because the layer it sits on is the layer we build. File Paperclip issues with them, not us.

The idea

Most AI tools ask you to manage prompts. Paperclip asks you to manage an organization. You set a goal at the top, hire agents into specific roles, and the work flows through the hierarchy without you in the middle of every hand-off.

The framing is not a metaphor. Roles, managers, budgets, tasks, approvals, and audit trails are the actual data model.

That distinction matters. A single agent can be impressive in a narrow session. A group of agents becomes hard to reason about quickly: who owns the goal, who is allowed to spend money, who can approve a destructive action, and where did the decision come from? Paperclip gives those questions a first-class place to live. It’s the org shape this enables — humans on the decisions, agents on the execution, structure visible as system state instead of buried in transcripts.

What you get

  • Roles and reporting lines. A goal-owning agent delegates to specialists instead of dumping everything into one giant prompt. Every hand-off is a logged object you can audit later.
  • Per-agent budgets. Each agent has a spend cap. It stops when it hits the cap. No surprise invoices, no 3am token bonfire.
  • You are the board. Approvals, overrides, pause, replace. Costly or destructive moves ask for sign-off by default.
  • Schedules instead of cron. Agents wake on a cadence, check their queue, act, and sleep. Daily standups, weekly reports, monthly cleanup.
  • Every action is a ticket. Tasks persist. Conversations are threaded. Decisions are logged. The trail is queryable.

The point is not to make agents look like employees for fun. The point is to make delegation observable. If an agent hires another agent, spends a budget, changes direction, or asks for approval, that should be visible as system state, not buried in a transcript.

Who it is for

  • Founders standing up an AI-first company who do not want to hire fifty people to find out the idea works.
  • Operators and agencies running multiple businesses from one control plane, with data isolation per tenant.
  • Anyone who has caught themselves babysitting AI tools and thought, “this is the part I was supposed to be automating.”

It is especially useful once the question changes from “can one model do this task?” to “can I keep a small AI team pointed at this outcome for weeks without losing the plot?”

When not to bother

  • You only ever run one agent. For solo use, a plain 5dive agent or CLI session is simpler.
  • You want a chatbot. That is not what this is.
  • You want a drag-and-drop node graph. Also not what this is.

Paperclip is overkill for most teams

Most 5dive users don’t run Paperclip, and don’t need to. The runtime’s own messaging — 5dive agent send and 5dive agent ask — already covers the coordination most teams actually do: one agent delegates a task, another waits on the reply, a third runs in parallel. Agents can pair with Telegram, share auth profiles, scope themselves to an isolation tier, hold a long-running tmux session through reboots, and call each other across the host without any orchestrator in the middle. That’s a small AI team you can build today, with no governance ceremony on top of it.

The honest take: Paperclip is structured-coordination overhead, and most workloads don’t need it. Roles, budgets, approval gates, audit trails as system state — those are valuable when you’re running an actual AI-staffed company with spend authority and destructive-action policies. For everything else, which is most things, 5dive alone is the simpler, smaller, faster path. Treat Paperclip as the optional upgrade, not the default — and don’t add it until you can name a thing it would solve that send and ask don’t already.

How they fit together (if you want both)

Paperclip is the workflow. 5dive is the bus.

Every agent on a 5dive host has two messaging verbs available to it — 5dive agent send (fire-and-forget) and 5dive agent ask (wait for the reply). They work between agents on the same machine without an orchestrator service in the middle. The CLI itself is the bus.

Paperclip can ride that bus. A Paperclip role assigned to “code reviewer” doesn’t need to wire up a custom RPC — it can shell out:

5dive agent ask code-reviewer "review PR #482, return verdict + 3 bullets" --timeout=300

Same pattern for the “research analyst,” the “deploy operator,” any specialist. Paperclip handles the org chart, the budget caps, the approval ceremony, the audit log. 5dive handles the underlying process, the auth, the isolation tier, the channel. Each layer is replaceable on its own without rewriting the other.

You can also run Paperclip without 5dive (it has its own self-host instructions). You can run 5dive without Paperclip (most of our users do, today). The two compose; neither is locked into the other.

And if you do want both: one click

Paperclip is open source and self-hosted. You own the data; there is no Paperclip account to sign up for. A web search for “Paperclip AI” will get you to the project and its docs.

What Paperclip still needs is a place to run: a server, a database, credentials, HTTPS, process supervision, updates, and backups. That is where 5dive comes in — first as a fully functional agent platform in its own right, and as a bonus, with a one-click Connect Paperclip flow in the dashboard for users who want the governance layer on top. It runs the installer (npm install -g paperclipai), provisions Postgres, wires Caddy and systemd, and hands you the URL. The org chart is running on your own machine in a few minutes — no manual server plumbing, no surprise dependencies. Updates, snapshots, and HTTPS are the platform’s job from then on.

Same software, less ops, no extra step. We are biased, obviously: 5dive is us. But the honest pitch is that 5dive on its own runs perfectly well as your agent platform; Paperclip is an optional upgrade for teams that specifically want the structured-coordination layer, and the connect flow is there so it’s a non-event when you do.