---
title: OpenClaw
description: Run OpenClaw, a personal AI assistant that answers in your chat apps and its own Control UI, with its memory on a volume.
section: Templates
order: 4
updated: 2026-11-04
---

[OpenClaw](https://openclaw.ai) is a personal AI assistant you run yourself, which you talk to from Telegram, Discord or its web Control UI. The **OpenClaw** template runs its gateway from the `ghcr.io/openclaw/openclaw:latest` image.

## What the template creates

| Resource | What it is |
| --- | --- |
| `openclaw` | The gateway and Control UI on port 18789, with 1 GiB of memory and a 1 GB volume at `/data` |
| `OPENCLAW_GATEWAY_TOKEN` | A token Dekka generates, which you sign in to the Control UI with |

OpenClaw keeps its state in `/data/.openclaw` and its workspace in `/data/workspace`, so its memory and settings survive every deploy. Dekka checks its health at `/startupz`.

OpenClaw needs 1 GiB of memory, which the Growth plan allows.

## What you provide

Every input is optional, and you can add or change any of them later in the service's variables.

| Input | What it's for |
| --- | --- |
| Anthropic API key | The model OpenClaw thinks with. You can add other providers later in its settings. |
| OpenAI API key | If you'd rather use OpenAI models. |
| Telegram bot token | To talk to OpenClaw from Telegram. Create a bot with @BotFather to get one. |

OpenClaw needs at least one model provider to answer anything, so add the Anthropic or OpenAI key now or in its settings before you start.

## After it deploys

1. Open the **Environment** tab, choose `openclaw`, and press and hold `OPENCLAW_GATEWAY_TOKEN` to reveal it.
2. Open the service's URL and sign in to the Control UI with that token.
3. Add model providers and chat channels in its settings, and if you added a Telegram token, send your bot a message.

The gateway token is a secret like any other, so treat it as a password. Anyone with it can use your assistant and the provider keys behind it.

## Limits

The volume means the service runs a single replica. During the pilot, deploys run on a simulated engine, so OpenClaw doesn't actually start or answer messages.

OpenClaw is open source under the [MIT licence](https://github.com/openclaw/openclaw). Your model provider's terms and charges apply to the requests it makes.