MCP for Claude, Cursor and ChatGPT
Bunker speaks the Model Context Protocol, so agents can create apps, deploy files, edit, share and roll back, with your permissions and nothing more.
Remote MCP: just the URL
Section titled “Remote MCP: just the URL”Add a connector (Claude, ChatGPT) or an MCP server (Cursor, VS Code, Claude Code) with this URL:
https://bunker.example.com/mcpThe client discovers everything else:
- Its first request gets
401withWWW-Authenticate: Bearer resource_metadata=".../.well-known/oauth-protected-resource/mcp". - It reads the protected-resource metadata (
/.well-known/oauth-protected-resourceand/.well-known/oauth-protected-resource/mcp) and the authorization-server metadata (/.well-known/oauth-authorization-server, also at/.well-known/openid-configuration). - It identifies itself with a client ID metadata document (an https URL as
client_id), or registers itself at/oauth/register. - Your browser opens Bunker’s consent page (
/oauth/authorize). You sign in as usual (built-in account or company SSO) and see which app is asking and where it will send you back. Plain-http redirect URIs are only accepted forlocalhost/127.0.0.1. - The client exchanges the code at
/oauth/token(PKCES256only, bound to the/mcpresource). Access tokens last 1 hour (BUNKER_OAUTH_ACCESS_TTL, in seconds); refresh tokens rotate on every use, and reusing an old one revokes the whole family. Revoke at/oauth/revokeor from your token list.
{ "mcpServers": { "bunker": { "url": "https://bunker.example.com/mcp" } } }claude mcp add --transport http bunker https://bunker.example.com/mcpBunker speaks the stateless 2026-07-28 protocol and the earlier session-based versions (2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05), so new and older clients both work.
Local MCP: stdio
Section titled “Local MCP: stdio”For agents that run on your machine, using your CLI login:
bunker login https://bunker.example.comclaude mcp add bunker -- bunker mcp{ "mcpServers": { "bunker": { "command": "bunker", "args": ["mcp"] } } }| Tool | |
|---|---|
create_site_from_html |
Publish an HTML page as a new app |
deploy_files |
Deploy files to an app (publish or draft) |
edit_file |
Edit one file of an app |
build_with_ai |
Ask Bunker’s AI builder to build or change an app |
get_sdk_docs |
The SDK, sandbox and deploy rules |
| One tool per API operation | Generated from GET /api/v1/ops: list, share, versions, restore, data, stats… |
fn_list, fn_call, fn_runs |
List an app’s server functions, call one, and read recent runs with their logs (handy for an agent fixing its own function) |
secrets_list, secrets_set, secrets_delete |
Manage an app’s function secrets. Values can be set, never read |
Destructive operations (delete, unshare, discard) are marked so clients can ask you first.
Resources: bunker://docs/agent-guide, bunker://docs/sdk, bunker://docs/sandbox, and bunker://sites/<app> for each app you can open.
Safety
Section titled “Safety”- The agent acts as you. It can’t see apps you can’t see.
- Page-only limits still apply: an app’s code can never deploy or share; only you, through the agent, can. The same goes for server functions.
- An agent can store a secret you give it, but can never read one back.
- Every action is in the audit log, marked as coming from a token.