Connecting OpenClaw

OpenClaw is a code-running CLI coding agent. Unlike Claude Code, Cursor, or Codex, it does not speak MCP natively — but it can run shell commands, which is all it needs to drive AgentPort through the ap CLI.

This page walks through having OpenClaw set itself up. You generate an API key, paste a short prompt into OpenClaw, and OpenClaw handles the rest — installing the CLI, authenticating it, and pulling in the agent skills that teach it the gateway's conventions.

Prerequisites

OpenClaw installed and working in your shell.
An AgentPort account, either on the cloud or your own self-hosted instance.
An AgentPort API key (see Step 1).

Step 1: Generate an API key

In the AgentPort UI, open the Developer page and scroll to API Keys. Generate a new key and copy it somewhere safe.

API keys are the right credential for an unattended agent like OpenClaw — they let it install integrations and call tools, but not change approval policies, view logs, or anything else administrative. The agent gets exactly the surface it needs.

Step 2: Paste this prompt into OpenClaw

Start an OpenClaw session and paste the following, replacing <your-api-key> with the key from Step 1:

Let's install AgentPort for managing third-party integrations.

1. Install the AgentPort CLI from npm with `npm install -g agentport-cli`
2. Authenticate it by running `ap auth login --api-key <your-api-key>`
3. Add the AgentPort skills from https://github.com/yakkomajuri/agentport-skills

Each step is something OpenClaw can do entirely on its own from a shell:

Install the CLI. npm install -g agentport-cli puts the ap (and agentport) binaries on your PATH. Full reference: CLI.
Authenticate. ap auth login --api-key … writes the key into ~/.config/agent-port/config.json so subsequent ap calls are authorized. There's no browser flow because there doesn't need to be — the key is already in OpenClaw's hands.
Add the skills. The agentport-skills plugin teaches OpenClaw the parts of the gateway that aren't obvious from the CLI's --help text — discovery order, how to handle the approval-required exit code, what to put in --info, and so on. The relevant skill for OpenClaw is agentport-cli.

What you can do next

In the same OpenClaw session, ask:

List my AgentPort integrations.

OpenClaw should run ap integrations list --installed -o json and report back what you have connected. If the list is empty, install something from the Integrations page in the AgentPort UI and try again. From here, OpenClaw can describe tools, call them, and surface approval URLs to you when a call is gated.

How OpenClaw uses AgentPort

OpenClaw drives AgentPort by shelling out to the ap CLI: ap auth status to confirm it's logged in, ap integrations list --installed to find what it can use, ap tools describe to read a tool's input schema, and ap tools call --integration <id> --tool <name> --args '<json>' --info '<why>' -o json to actually execute a tool.

The agentport-cli skill is what makes this work end-to-end. It pins down the exit-code contract — 0 for success, 2 plus Approval required: <url> on stderr when a tool needs human sign-off, 1 for a hard denial — so OpenClaw can branch on outcomes correctly. When it hits an approval gate, the skill tells it to surface the URL to you in chat and retry the identical command once the approval lands.

See the CLI reference for every command and flag, and the Skills page for the full bundle.

Troubleshooting

ap: command not found after install. The npm global install directory isn't on your PATH. Run npm config get prefix to find it; the binaries live in <prefix>/bin. Add that to your shell's PATH and start a new OpenClaw session so it picks up the change.
ap auth login fails or ap auth status shows auth_mode: none. Re-check the API key — it should start with ap_ and be pasted in full with no surrounding whitespace. If you're on a self-hosted instance, point the CLI at it first with ap auth set-instance-url https://<your_domain> (this clears any previous credentials), then re-run ap auth login --api-key ….