Docs /

AI agents

Coding agents (Claude Code, Cursor, Codex and others) deploy with the same CLI as people, and chat assistants without a filesystem deploy through a remote MCP server. This page shows how to set each one up.

npx -y lobsite --yes

How agent mode works

The CLI knows it runs under an agent from the variables agents set: CLAUDECODE (Claude Code), CURSOR_AGENT (Cursor), CODEX_SANDBOX (Codex), GEMINI_CLI, COPILOT_AGENT, GOOSE_TERMINAL, REPLIT_AGENT and AGENT. Set LOBSITE_AGENT=<name> to name another agent, or LOBSITE_AGENT=0 to turn agent mode off. Under an agent:

• Nothing is ever asked. The first deploy of a folder prints its summary and carries on; commands that can't be undone need --yes.
• There are no colours, spinners, QR codes or clipboard copies, only plain lines.
• A deploy ends with a line written for the agent to pass on:

Tell the user: the site is live at https://quiet-fox-4821.lobsite.app. It's temporary and is removed in 6h. To keep it, open https://lobsite.com/claim/quiet-fox-4821#lob_claim_… and sign in (free).

Sites deployed without an account last 6 hours. The claim link keeps the site: the person opens it and signs in, without touching a terminal. Running the same command again updates the same site, because lobsite.toml remembers it.

Claude Code

Claude Code can deploy with the shell command alone. To have it use Lobsite every time, add a line to the project's CLAUDE.md:

Deploy with `npx -y lobsite --yes` and relay the "Tell the user:" line from the output.

Or add the MCP server, which gives Claude tools to deploy, list sites and deploys, roll back and promote:

claude mcp add lobsite -- npx -y lobsite@latest mcp

Cursor

Add the MCP server to .cursor/mcp.json:

{ "mcpServers": { "lobsite": { "command": "npx", "args": ["-y", "lobsite@latest", "mcp"] } } }

Or add a project rule in .cursor/rules/ with the section from AGENTS.md.

Codex and other agents

Codex and the other agents that read AGENTS.md pick up a section like this one. Paste it into the project's AGENTS.md:

## Deploying

This site deploys with [Lobsite](https://lobsite.com). Build it, then from the project folder run:

```bash
npx -y lobsite --yes
```

- Tell the user the URL, and relay the "Tell the user:" line from the output word for word.
- Sites deployed without logging in are removed after 6 hours. The claim link in the output keeps the site; the user opens it and signs in.
- `lobsite.toml` remembers the site, so the same command updates it. Use `--preview` to deploy without changing the live site, and `--json` for machine-readable output.
- If a deploy fails, follow the fix in the error message. Never pass `--include-secrets`.

Any agent that speaks MCP over stdin/stdout can run npx -y lobsite@latest mcp instead.

ChatGPT and Claude.ai

Chat assistants write pages but have no folder to deploy. Add Lobsite as a custom connector (an MCP server) with this URL:

https://lobsite.com/api/mcp

It speaks Streamable HTTP and has one tool, deploy_files, which takes the files' contents. Then ask the assistant to put the page online. Sites deployed this way have no account: the answer carries the claim link that keeps them.

Without Node: curl

Where npx can't run, one curl command deploys a page or a few files and answers with the URL, the claim link and how to update the site:

curl --data-binary @index.html https://lobsite.com/api/v1/drop
curl -F index.html=@index.html -F css/site.css=@site.css https://lobsite.com/api/v1/drop

To update the same temporary site, pass its name and the claim token from the first answer:

curl --data-binary @index.html -H "Authorization: Bearer lob_claim_…" "https://lobsite.com/api/v1/drop?site=quiet-fox-4821"

?preview=1 deploys without changing the live site. A direct deploy carries up to 25 MB. Add ?format=json for a JSON answer.

MCP tools

The local tools run the CLI commands with --json --yes and take path, the project folder (by default, the folder the server started in). deploy answers with the "Tell the user" line first. Logging in and deleting sites aren't tools: one needs the person in a browser, the other can't be undone.

JSON output

--json prints one JSON object on stdout, with logs on stderr:

{
  "site": "quiet-fox-4821",
  "deploy_id": "d7k2m9qa",
  "url": "https://quiet-fox-4821.lobsite.app",
  "preview_url": "https://d7k2m9qa--quiet-fox-4821.lobsite.app",
  "live": true,
  "expires_at": "2026-10-13T16:20:38Z",
  "claim_url": "https://lobsite.com/claim/quiet-fox-4821#lob_claim_…",
  "files": { "total": 42, "uploaded": 3 },
  "bytes_uploaded": 14336,
  "duration_ms": 812
}

claim_url and expires_at are null for sites with an account. A failure is { "error": { "code", "message", "why", "fix", "exit_code", "retryable" } }. Branch on code, which never changes (see Errors), and retry only when retryable is true.