> ## Documentation Index > Fetch the complete documentation index at: https://getoverlay.io/docs/llms.txt > Use this file to discover all available pages before exploring further. # Claude Code and Codex agents > Run Claude Code or Codex as an agent on an Overlay Cloud computer, bring your own setup, and give it your Overlay workspace. You can run **Claude Code** or **Codex** as a teammate on its own computer. Overlay starts the computer, keeps the agent signed in with your account, and gives it your workspace as tools. Overlay Cloud agents need a paid plan. ## Create one 1. **Agents → New agent → Other agent**. 2. Choose **Claude Code** or **Codex**, and **Overlay Cloud**. 3. Choose the account it runs on. If you have none, the dialog offers **Connect** right there, and the account is also saved under Settings → Agents → Agent accounts: * **Claude Code**: paste a token from running `claude setup-token` on your own computer (for a subscription), or an Anthropic API key. * **Codex**: sign in with ChatGPT (Settings → Agents shows a code and a link to enter it), or an OpenAI API key. * **OpenCode, Cursor** (experimental): an OpenRouter key (OpenCode) or a Cursor key. Secrets are stored encrypted and are never shown again. 4. Choose a model (Claude Code and Codex): leave it on Default, or pick one. Codex also has a Reasoning level. You can change it later on the agent page; it applies from the next message. 5. Choose a size: Small (2 vCPU, 4 GB), Default (4 vCPU, 8 GB), or Large (8 vCPU, 16 GB). 6. Create. The dialog shows the steps (allocating, booting, connecting) and opens the conversation when the agent is ready, usually within a minute. You need enough Overlay credit to run the computer (at least \$1); without it Overlay will not start one. ## The computer * It **pauses itself after about 10 minutes idle** and wakes when you message the agent. A wake can take up to a minute or two; if the message says the environment went offline, send it again. * On the agent's page: **Pause**, **Resume**, **Restart**, and the size and account it uses. **Restart** also recovers a computer that stopped answering. Do not pause or restart while the agent is replying: the interrupted answer is stopped and you will need to send it again. * **Needs sign-in** means the account was rejected. Use **Reconnect** and enter a fresh token. * **Unavailable** means the computer is gone, usually because credit ran low. Add credit, then archive the agent and create it again. * Compute is charged to your Overlay credit while the computer runs. Model use on your own subscription or key is not charged by Overlay. ## Give it your Overlay workspace The page's **Overlay access** control decides what the agent can do in Overlay: **None**, **Read only**, **Read and write**, or **Everything** (the default). The agent reaches Overlay through its MCP server and acts as whoever messages it. See [Tools](https://getoverlay.io/docs/help/tools.md). ## Bring your setup (Config) The **Config** section copies your Claude Code or Codex configuration onto the computer: instructions (`CLAUDE.md` or `AGENTS.md`), skills, commands, subagents, output styles, prompts, settings, and MCP servers. **What is never copied**: sign-in credentials, history, plugins, and caches. Secret-looking values (an MCP server's token) are replaced by names, and you enter each value once; it is stored encrypted and given to the agent only while it runs. Hooks and notify or status commands are listed and stay **off** until you turn them on. Permission modes that skip or deny Overlay's approval prompts (`bypassPermissions`, `acceptEdits`, `dontAsk`) are not imported. Two ways in: * **From your computer**: choose **Import from your computer**, copy the command, and run it in a terminal on the computer that has your setup. It needs Node 24 (the command fetches it). It lists what it found, and `--dry-run` sends nothing. * **Upload a folder**: choose **Upload .claude folder** (or **.codex**) and pick the folder. Then review: counts of what was found, what was left out and why, the values it needs, and any hooks. **Apply to the machine** writes it. Each import is a version; **Earlier versions → Restore** goes back, removing what a later version added. The agent reads skills and settings when a session starts, so changes show on its next message. ## Agents of your own Both kinds can ask each other for help. See [Agents](https://getoverlay.io/docs/help/agents.md). ## If something goes wrong | You see | What to do | | - | - | | "could not start this turn" | The computer may be waking or a previous answer was interrupted. Wait a minute and send again. | | "The environment went offline" | The computer was asleep. Send the message again. | | "Needs sign-in" | Reconnect the account on the agent's page. | | The agent shows "Unavailable" | The computer was deleted (often low credit). Add credit and recreate the agent. | ## How replies read An agent's reply shows its final answer, with its tool calls and the text it wrote between them folded behind **Worked for…**; click it to see everything in order. Agents are asked to end with a short summary of what they did. Turn this off in Settings → Agents (**Collapse working details in responses**) to show the whole response. Chats with Overlay's own models are not affected. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.