Skip to main content

Quickstart

From nothing to your first model call, in six steps.

Steps 1 to 5 are done once, by an admin. Step 6 is done by everyone who calls the gateway.

The console tracks these steps too

The AI Gateway Overview shows the same setup as a checklist under Get started, marks each step as you finish it, and links to where each one is done.

1. Switch identity on for the project​

The gateway checks every token against an identity service that runs in your project. It has to be running first.

Go to Account Settings → Identity, open the Dataplane IDP tab, find your project and choose Enable.

→ Identity setup

2. Turn the gateway on​

Open AI Gateway — the first item under Serve — and choose Enable AI Gateway on the Overview. It provisions into your data plane in about 4 minutes and comes up Live, with your base URL.

On AWS, your customer stack must also have the AI Gateway credential vault turned on. If it does not, the button says so and links to the fix.

→ Enable the gateway

3. Connect a model provider​

Open Models & Providers, then the Providers tab. Choose Add provider, pick Anthropic, OpenAI, Baseten or AWS Bedrock, and pass Test connection before you save.

→ Connect a provider

4. Publish a model​

On the Models tab, choose Add a model. Pick which provider's model serves it, and give it an Alias name your clients will use — for example claude-opus-5.

→ Publish a model

5. Grant access to it​

A new model reaches nobody until it is granted. Open the model's Permissions tab and grant Invoke to yourself, or to whoever will call it.

→ Permissions

6. Sign in and call it​

Install the oh CLI, sign in once, and start your client through it:

oh ai-gateway login --gateway-url https://ai-gateway-<id>.dp.onehouse.ai
oh ai-gateway run -- claude

The CLI keeps your token fresh for the whole session. See Sign-in and tokens.

Or, for a quick test: on the Overview, under Call the gateway, choose Get my token and sign in. Copy your token, then:

curl https://ai-gateway-<id>.dp.onehouse.ai/v1/chat/completions \
  -H "X-Onehouse-Authorization: Bearer $ONEHOUSE_GATEWAY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "messages": [{"role": "user", "content": "Say hello."}]
  }'

A completion comes back. You are done.

Setting up an agent instead of a person? See Agent access.

Already using an OpenAI-compatible client?

Change the base URL, and send the token in the X-Onehouse-Authorization header. Nothing else. See Connect your client for Claude Code, Codex, VS Code and JetBrains.

Optional: give agents tools​

Register MCP servers on the MCP Servers tab, expose the tools you want, and grant Call on them. For an OAuth server, also grant Connect on the server so people can sign in to it. See Tool servers.

On AWS, the Onehouse MCP server is already registered.