Paperclip AI + OpenClaw: Integration & Gateway Adapter Setup Guide

The Paperclip AI OpenClaw integration is the most common way people build their first agent company. OpenClaw gives you a capable personal assistant that already talks to WhatsApp, Telegram, Slack and Discord. Paperclip gives that assistant a job title, a manager, a budget and a task queue. This guide explains how the OpenClaw gateway adapter works, walks through the setup, and covers the pitfalls we see when businesses wire the two together. If you are new to the platform, start with our complete Paperclip AI guide and come back here once you have a running instance.

What OpenClaw is, and what it is not

OpenClaw is an open-source, Node.js based personal assistant gateway. You run it on a server or a laptop, connect it to a model provider, and it becomes reachable through more than two dozen messaging channels. The big four are WhatsApp, Telegram, Slack and Discord, but the community has added channels for email, SMS and voice. It has the largest skills ecosystem of the 2026 open-source agents, which is why so many people already have one running.

What OpenClaw does not have is any concept of an organization. One instance is one assistant. It has no idea that another OpenClaw instance exists, no shared goal, no way to hand off work, and no spending cap other than what your model provider enforces. That gap is the entire reason the Paperclip project exists. The Paperclip README puts it in one sentence: if OpenClaw is an employee, Paperclip is the company.

How the OpenClaw gateway adapter works

Paperclip does not run agents itself. It is a control plane. Every agent is attached to Paperclip through an adapter, and the adapter's only real job is to deliver a heartbeat and collect the result. The OpenClaw gateway adapter follows this pattern:

1. Paperclip fires a heartbeat. Heartbeats run on a schedule you set per agent, or fire immediately when a ticket is assigned to that agent.
2. The adapter calls the OpenClaw gateway. It sends the agent's identity, the checked-out ticket, the goal chain that ticket belongs to, and any skills the role is allowed to use.
3. OpenClaw does the work. It uses its own tools, memory and channels exactly as it would if a human had messaged it.
4. The adapter reports back. Tool calls, token usage and the final output are written to the ticket, the cost is charged against the agent's monthly budget, and the ticket moves to review or done.

Because the ticket is checked out atomically, two OpenClaw employees can never grab the same piece of work. If the instance crashes mid-task, the ticket stays checked out with a timestamp and payload, and Paperclip can reassign it on the next heartbeat. That behavior alone fixes the biggest complaint people have about running several assistants by hand.

Step-by-step setup

The exact field names in the Paperclip UI change between releases (the current one is v2026.916.1 as of September 2026), so this is written at the level of what you are doing rather than which box you click. Check the adapter page in the official docs for the current form.

1. Get Paperclip running

npx paperclipai onboard --yes

That starts the server with an embedded PostgreSQL database and opens the dashboard at http://localhost:3100. If you want a durable server install instead, follow our Paperclip AI install guide.

2. Run OpenClaw and expose its gateway

Your OpenClaw instance needs to be reachable from the Paperclip server. On a single machine that is a localhost port. Across machines, put both on a private network (Tailscale is the common choice) rather than opening the gateway to the public internet. Generate a gateway token in OpenClaw and keep it; Paperclip will store it in its encrypted secrets vault.

3. Create a company and a role

In Paperclip, create a company, then add a role to the org chart. Give it a real job description. The job description becomes part of every heartbeat, so a vague one produces vague work. Something like "Customer Support Specialist: answers inbound Telegram questions about order status, escalates refunds over $200 to the Operations Manager" is far better than "support agent".

4. Add the agent with the OpenClaw adapter

Create an agent under that role, choose OpenClaw as the adapter, and enter the gateway URL and token. Set a heartbeat schedule. For a reactive support role, a short interval or task-triggered heartbeats make sense. For a research role that batches work, once an hour is plenty.

5. Set a budget

Assign a monthly USD limit. Paperclip tracks cost per agent and pauses the agent automatically when it hits 100 percent of the limit. Start low. A $25 cap on a new agent tells you in a week what a month of real usage will cost.

6. Send a test heartbeat

Create a small ticket, assign it, and trigger a heartbeat manually. Open the ticket's activity log. You should see the adapter call, the tool calls OpenClaw made, the token count and the cost. If the log is empty, the gateway URL or token is wrong, or the network path is blocked.

Running multiple OpenClaw employees

The interesting part starts when you add a second and third instance. Each one is a separate OpenClaw process with its own memory, its own channels and its own token. In Paperclip they become separate employees with separate roles and budgets, reporting to a manager agent or to you as the board.

Role	OpenClaw channels	Heartbeat	Budget	Reports to
Support Specialist	Telegram, WhatsApp	Task-triggered	$40/mo	Operations Manager
Community Manager	Discord	Every 30 min	$30/mo	Marketing Lead
Research Analyst	Slack (internal)	Hourly	$60/mo	Marketing Lead
Marketing Lead	Slack (internal)	Every 2 hours	$50/mo	Board (you)

The manager roles do not have to be OpenClaw. Many teams use a Claude Code or Codex agent as the manager because those harnesses are better at reading a large backlog and delegating, and keep OpenClaw for the roles that need messaging channels. Our post on Paperclip AI and Claude Code covers that pattern.

Pitfalls we see repeatedly

• Sharing one OpenClaw instance across several Paperclip agents. It works technically, but the instance's memory bleeds between roles and the audit trail becomes meaningless. One instance per employee.
• No budget on day one. An OpenClaw agent with a heartbeat every five minutes and no cap can burn through a model provider balance overnight. Set the cap before the first heartbeat.
• Public gateway exposure. The gateway is an authenticated endpoint, but it should still sit on a private network. Treat the token like a database password.
• Skipping the job description. Paperclip's goal alignment only works if the role knows what it is for. Write the job description the way you would for a human hire.
• Ignoring the approval gates. Anything that sends a message to a customer, spends money or changes production data should route through an approval stage until you trust the role.

When you would choose Hermes instead

OpenClaw is the right pick when the job is about channels: talking to customers, moderating a community, answering internal questions in Slack. Hermes Agent, from Nous Research, is the right pick when the job is about getting better over time. Hermes builds reusable skills from its own experience, which makes it a stronger fit for repetitive back-office work that evolves. Plenty of companies run both under one Paperclip org. We explain the tradeoffs in Paperclip AI and Hermes Agent and in the head-to-head Paperclip AI vs OpenClaw comparison.

Frequently asked questions

Do I need OpenClaw to use Paperclip AI?

No. Paperclip supports Claude Code, Codex, Cursor, Hermes, OpenCode, Pi, Gemini and plain HTTP agents. OpenClaw is popular because it is the easiest way to get a messaging-connected employee, but it is one adapter among many.

Can Paperclip and OpenClaw run on different servers?

Yes. The adapter only needs network access to the OpenClaw gateway. Put both on a private network such as Tailscale and point the adapter at the private address.

Does Paperclip replace OpenClaw's own scheduling and memory?

No. OpenClaw keeps its memory and skills. Paperclip adds the layer above it: tickets, goals, budgets, approvals and the audit log. Its heartbeat scheduler decides when the agent works, and OpenClaw decides how.

How much does the integration cost?

Both projects are open source and free. Your costs are the server and the model provider usage each OpenClaw instance generates, which Paperclip tracks per agent. See our Paperclip AI pricing breakdown.

Can one OpenClaw agent delegate to another?

Not directly. Delegation happens through Paperclip. A manager agent creates a ticket and assigns it to a subordinate role, and the subordinate picks it up on its next heartbeat.