Reference

Connect an MCP-native agent

Attach a Meltbox workspace to any MCP-native agent. Claude and ChatGPT connect with just the URL. OAuth handles the rest, and in ChatGPT your briefs render as interactive cards. Header-capable clients (Cursor, VS Code, Windsurf, Cline, Claude Code) attach with the same URL and a scoped bearer token. Either way the workspace shows up as a set of tools that push briefs, read the reply, discover integrations, and broker secrets through the proxy.

The endpoint

One URL is the whole server. It speaks JSON-RPC 2.0 over a single POST:

Authentication

Two ways in. Both end as the same kind of scoped, revocable credential on the Access page (API keys):

A missing or invalid credential returns 401 with a WWW-Authenticate header pointing at the OAuth discovery metadata (RFC 9728). OAuth-native clients take it from there. Dynamic client registration (RFC 7591), PKCE S256, refresh, and revocation (RFC 7009) are all supported.

Mint a connector token

In the app, open Agents → MCP connector. Name the connection and choose its access (Briefs-only or Full), and Meltbox mints a scoped token and shows a paste-ready config for your client. The connection then appears under Connected agents, where you can revoke it anytime.

Already redeemed a curl invite link? You don't need a second credential. The redeem response includes an mcp block { url, transport, header }. Attach that same token as the header below.

Set up your client

ChatGPT · Claude: connect by URL (OAuth)

1. ChatGPT: Settings → Apps & connectors → Advanced settings → enable Developer mode, then create a connector with MCP server URL https://meltbox.ai/mcp and authentication OAuth. Developer mode is only needed until the Meltbox listing lands in the ChatGPT app directory.
2. Claude (web / desktop / mobile): Settings → Connectors → Add custom connector → paste https://meltbox.ai/mcp.

Either way you land on the Meltbox consent page: sign in, pick the workspace, approve. The connection shows up on the Tokens page (via oauth), revocable anytime. In ChatGPT, Meltbox is a full ChatGPT app. Briefs render as interactive cards in the conversation.

Header clients

Clients that let you set a custom Authorization header attach directly with a minted token.

Claude Code (CLI): native header

claude mcp add --transport http meltbox https://meltbox.ai/mcp \
  --header "Authorization: Bearer mb_your_token"

Cursor · Windsurf: native header (e.g. ~/.cursor/mcp.json; VS Code uses a servers key)

{
  "mcpServers": {
    "meltbox": {
      "url": "https://meltbox.ai/mcp",
      "headers": { "Authorization": "Bearer mb_your_token" }
    }
  }
}

Claude Desktop: via mcp-remote (claude_desktop_config.json)

{
  "mcpServers": {
    "meltbox": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://meltbox.ai/mcp", "--header", "Authorization:Bearer mb_your_token"]
    }
  }
}

The tools

Attaching the server exposes nine tools. Call whoami first; the rest are the brief loop, both sides of it, plus resource discovery and the proxy.

The loop

The core Meltbox loop is the same over MCP as everywhere else. Orient, say hello, then watch for the human's reply:

1. Orient. Call whoami to confirm which workspace you're connected to, your role, and your scopes.
2. Say hello. Call push_brief with something substantive (a plan, a set of options, a draft) as a clickable page with a notes box. It lands in the operator's inbox and returns {id, url, deeplink}.
3. Watch. Call poll_brief_events repeatedly with the cursor it returns, reading the operator's selections and notes as they arrive and acting on them. A status: archived event means they've reviewed it and made their decisions. Act on the feedback and push the next round as a new brief.

To do real work alongside the loop, discover the workspace's integrations with list_resources, ask for access you don't have with request_scope, and make brokered reads with proxy. The secret stays server-side and never reaches you. See Authoring briefs for how to build a good first brief.

The ChatGPT app

Connected to ChatGPT, Meltbox is more than a toolbox. It's an app. Ask "what's in my Meltbox inbox?" and the triage list renders in the conversation. Open a brief and you see the actual page your agent authored: its choice buttons record real selections, the note box feeds the agent, archive closes the round. The same loop you'd run in the Meltbox app, without leaving the chat.

• Operator side: list_briefs → get_brief → respond_to_brief, by clicking the card or just saying the decision.
• Agent side. ChatGPT can also produce work: push_brief delivers to your inbox and shows a preview card of what it sent.

Where to next