MCP Protocol Reference

CiteKit implements the Model Context Protocol (MCP) to allow AI agents to browse, understand, and extract multimodal evidence directly.

---

Tools

The CiteKit MCP server exposes the following tools to the agent:

listResources

Returns a list of all resource IDs currently available in the local CiteKit index.

• Input: {}
• Output: {"resources": ["lecture_vid", "book_pdf", ...]}

getStructure

Retrieves the full semantic map (JSON) of a specific resource.

• Input: {"resource_id": "string"}
• Output: The full ResourceMap schema.
• Agent Usage: Agents call this first to "see" what is inside a file without downloading it.

getNode

Fetches metadata for a specific node ID.

• Input: {"resource_id": "string", "node_id": "string"}
• Output: The single Node object.
• Agent Usage: Used for fast verification of a specific concept's location.

resolve

The "Resolution" tool. Converts a node into evidence.

• Input:
  • resource_id: (Required)
  • node_id: (Required)
  • virtual: (Optional, default: False)
• Output: ResolvedEvidence object (contains absolute path to clip/slice).
• Agent Usage: The final step. The agent takes the output_path and attaches it to the current chat context as grounded evidence.

---

Internal Protocol Flow

1. Discovery: User asks "What happens at 5 minutes in the video?".
2. Mapping: Agent looks at listResources.
3. Selection: Agent calls getStructure(lecture_vid).
4. Pinpointing: Agent finds a node with id: "recursion_demo" and location: {start: 300, end: 360}.
5. Resolution: Agent calls resolve(lecture_vid, "recursion_demo").
6. Grounding: CiteKit extracts the 60s clip, returns the path, and the agent "sees" the video.

---

Configuration

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "citekit": {
      "command": "python",
      "args": ["-m", "citekit.cli", "serve"]
    }
  }
}

Cursor / Cline / Roo Code

See the MCP Integration Guide for specific IDE setup.