MCP — Connect Caster to ChatGPT, Claude, and Gemini

Quick install: caster.zip/connect — copy the URL, tap a button, paste in the AI app. Two taps total.

Caster exposes a Model Context Protocol server so AI assistants — including the mobile apps for ChatGPT, Claude, and Gemini — can deploy and manage your sites directly from a chat. Open the AI app on your phone, ask it to publish something, and a real Caster instance appears.

Endpoint

https://mcp.caster.zip/

The MCP server speaks the Streamable HTTP transport (protocol version 2025-06-18) and is protected by OAuth 2.1 with PKCE. The first time you connect, the AI client opens a browser tab so you can sign in to Caster and approve the connector. After that, the client stores a Caster API key and calls the MCP transparently.

Tools

Tool What it does
deploy_instance Deploy a list of {name, content} files. Creates a new instance if instance is omitted, otherwise redeploys. Blocks up to 120 s waiting for the deploy to go live.
list_instances List all instances in your account.
get_instance Fetch instance details and recent deployments.
get_deployment_logs Read build logs for a specific deployment.

Tools are stable JSON-RPC: see the responses with curl https://mcp.caster.zip/ -X POST -H 'Content-Type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'.

Authentication

The MCP endpoint is OAuth-protected. Every tool call requires a Bearer token issued through the OAuth flow described above — your connector handles this once on connect and then transparently. Sites you deploy via deploy_instance land directly on your account and show up at caster.zip/console.

Need a one-shot anonymous deploy without an account (e.g. for a quick demo)? Use the REST endpoint instead:

curl -X POST https://caster.zip/api/v1/deploy \
  -H 'Content-Type: application/json' \
  -d '{"files":[{"name":"index.html","content":"<h1>hi</h1>"}]}'

It returns a claim_token you can later trade for ownership by signing in.

Connecting from each app

Claude (claude.ai web + Claude mobile)

  1. Open Settings → Connectors → Add custom connector.
  2. Paste https://mcp.caster.zip/ as the URL.
  3. Claude will redirect you to caster.zip to sign in. Approve the request.
  4. In any conversation, ask: "Publish a hello-world page on Caster."

ChatGPT (chatgpt.com + ChatGPT mobile)

  1. Settings → Connectors → Add MCP server.
  2. URL: https://mcp.caster.zip/. Auth: OAuth.
  3. ChatGPT opens caster.zip in a browser tab to sign in. Approve.
  4. Tools appear under the connector menu in chat.

Gemini

  1. Open the Extensions / Tools panel.
  2. Add a remote MCP server at https://mcp.caster.zip/.
  3. Complete the sign-in redirect on caster.zip.
  4. Ask Gemini to deploy a site — it'll call deploy_instance for you.

Revoking a connector

Each AI client gets its own Caster API key on approval. To revoke a single connector without affecting the others, go to Keys in the Caster console, find the row labelled OAuth: <client name>, and delete it. The next request from that client will fail; re-adding the connector will start a fresh OAuth flow.

OAuth metadata (for client developers)

Caster implements RFC 8414 (authorization-server metadata), RFC 9728 (protected-resource metadata), and RFC 7591 (dynamic client registration). The relevant URLs:

Document URL
Protected-resource metadata https://mcp.caster.zip/.well-known/oauth-protected-resource
Authorization-server metadata https://caster.zip/.well-known/oauth-authorization-server
Dynamic client registration https://caster.zip/oauth/register
Authorization endpoint https://caster.zip/oauth/authorize
Token endpoint https://caster.zip/oauth/token

Only PKCE (code_challenge_method=S256) is supported; no client secrets are issued. Access tokens are long-lived Bearer tokens; refresh tokens are not issued in v1, so clients re-run the flow when a token expires (default: 30 days).