@agent-format/mcp

@agent-format/mcp

MCP Apps server that renders .agent files as interactive dashboards inline in the chat of any MCP Apps–supporting client.

Status: Draft v0.1 — targets MCP Apps spec 2026-01-26.

What it does

When connected to Claude Desktop, ChatGPT (via Apps SDK), Cursor, VS Code Copilot, or Goose, this server exposes two tools:

• render_agent_file(path) — reads an .agent file from disk and renders it inline as a kanban / timeline / metrics / log / mindmap / etc. dashboard.
• render_agent_inline(data) — renders a full .agent JSON object that the agent just generated in this turn.

The rendered UI is the standard .agent viewer (knorq-ai.github.io/agent-format) embedded in the chat. All 12 section types work.

Install

npm install -g @agent-format/mcp

Or run without installing:

npx @agent-format/mcp

Configure your client

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on Windows:

{
  "mcpServers": {
    "agent-format": {
      "command": "npx",
      "args": ["-y", "@agent-format/mcp"]
    }
  }
}

Restart Claude Desktop. The first time a tool runs, Claude will ask for permission.

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "agent-format": {
      "command": "npx",
      "args": ["-y", "@agent-format/mcp"]
    }
  }
}

VS Code Copilot Chat

Add to settings.json:

{
  "mcp.servers": {
    "agent-format": {
      "command": "npx",
      "args": ["-y", "@agent-format/mcp"]
    }
  }
}

ChatGPT (Apps SDK)

ChatGPT Apps are registered via the developer portal, not a local config file. The stdio pattern above targets desktop clients. For ChatGPT, package this server as an HTTP endpoint — a future release will include that entrypoint.

Try it

After configuring a client and restarting:

"Render /Users/me/project.agent as a dashboard."

Claude calls render_agent_file, the viewer opens inline, and you see the kanban / timeline / metrics.

Or inline:

"Turn these TODOs into a kanban and render it."

Claude generates the .agent JSON, calls render_agent_inline with it, and the dashboard appears in the chat without touching disk.

How it works

• The server is a stdio-based MCP server using @modelcontextprotocol/sdk and @modelcontextprotocol/ext-apps.
• Tools declare a shared UI resource URI ui://agent-format/render.html.
• The UI resource is a single self-contained HTML document: the @agent-format/renderer React bundle and CSS are inlined at build time (see build-ui.mjs). No nested iframes, no external fetches, no CSP frameDomains required — the default sandbox is sufficient.
• When the tool result arrives via the MCP Apps ui/notifications/tool-result postMessage, the embedded script reads structuredContent.data and mounts <AgentRenderer/> against it directly.

This means the visual output is identical to what you'd see on the standalone viewer — same React renderer, same CSS — just embedded in the chat.

Development

cd packages/mcp
npm install
npm run build
node dist/server.js

To test end-to-end locally, point your client's command at your absolute dist/server.js path instead of npx.

License

MIT.