MCP Overview

Stackie’s MCP server gives AI agents full control over sandboxes, commands, logs, and health data through a structured JSON-RPC interface.

What is MCP?

The Model Context Protocol is an open standard for connecting AI agents to external tools and data sources. Stackie’s MCP server lets agents like Claude, GPT-4, and custom LLM clients control sandboxes, execute commands, stream logs, and query health data — all via a structured JSON-RPC interface.

Transport Modes

Stackie supports two transports:

  • SSE (Server-Sent Events) — HTTP-based streaming. Agents connect to http://localhost:<port>/sse and receive a stream of JSON-RPC messages. This is the default transport used by Claude Code and most web-capable agents.
  • stdio — Launch as a subprocess via stackie mcp serve --stdio. Used by agents that spawn MCP servers as child processes.

Tool Discovery

When an agent connects, it sends initialize followed by tools/list to discover available tools. No configuration is needed — agents discover capabilities automatically at connection time.

Authentication

The MCP server uses bearer tokens. Each agent gets a unique token at enrollment time (written to the agent’s config file). Tokens are stored as SHA-256 hashes in the database — raw tokens are never persisted.

Resource Providers

In addition to tools, the MCP server exposes resource URIs:

  • db://{sandbox_id}/{table} — Query sandbox databases
  • fs://{sandbox_id}/{path} — Read files inside sandboxes

Resources support both resources/read (single fetch) and resources/subscribe (live updates via SSE).