MCP Integration Overview
ChorumAI exposes your project memory to external AI agents through the Model Context Protocol (MCP).
Why This Matters
IDE agents like Claude Code, Cursor, and Windsurf are stateless by design—they start fresh every session. ChorumAI bridges this gap by giving these agents access to your persistent patterns, decisions, and invariants.
MCP vs Web UI
Not sure which to use? Here’s a quick comparison:
| Use Case | Web UI | MCP |
|---|---|---|
| Standalone chat with memory dashboard | ✅ | ❌ |
| Memory for Cursor/Claude Code/Windsurf | ⚠️ manual | ✅ automatic |
| Peer review across providers | ✅ | ❌ |
| Cost tracking & budgets | ✅ | ❌ |
| Review/approve learnings | ✅ | via web |
| Just want memory in existing tools | ❌ | ✅ |
Recommendation: Use both. MCP for day-to-day coding in your IDE, web UI for reviewing memory and managing learnings.
New to Chorum? Start with the 5-Minute Setup to get MCP running in your IDE.
What is MCP?
The Model Context Protocol is an open standard that allows AI assistants to connect with external tools and data sources. Think of it as a USB port for AI—a standardized way for different systems to talk to each other.
When you connect your IDE to ChorumAI via MCP:
- Your AI agent can query your project memory
- Relevant patterns and decisions are injected into context
- The agent can propose new learnings (pending your approval)
- Every interaction strengthens your project’s confidence score
Architecture
┌─────────────────┐ ┌─────────────────────────┐
│ Claude Code │ │ ChorumAI MCP │
│ Cursor │◄─── MCP Protocol ───►│ Server │
│ Windsurf │ (stdio/HTTP) │ │
│ Any MCP Client│ │ ┌─────────────────┐ │
└─────────────────┘ │ │ Project Memory │ │
│ │ Database │ │
│ └─────────────────┘ │
└─────────────────────────┘The MCP server runs locally via npx chorum-mcp. Your IDE communicates with it using the MCP protocol, and the server queries your ChorumAI memory on behalf of the agent.
Key Concepts
Read Operations (No Approval Required)
Your AI agent can freely read from your memory:
| Tool | What It Does |
|---|---|
chorum_list_projects | List all your projects |
chorum_query_memory | Semantic search through patterns, decisions, invariants |
chorum_get_invariants | Get all active “must not break” rules |
chorum_get_project_context | Get project metadata (tech stack, instructions) |
Write Operations (Human-in-the-Loop)
When an agent wants to add to your memory:
| Tool | What It Does |
|---|---|
chorum_propose_learning | Propose a new pattern or decision |
chorum_log_interaction | Log the interaction for confidence scoring |
Proposed learnings require your approval in the ChorumAI UI before they’re added to memory. This keeps you in control of what your AI “remembers.”
Security Model
- Token-based authentication — Each API token is scoped to a single user
- Revocable tokens — Revoke access anytime from Settings
- Human approval for writes — Agents can’t modify your memory without permission
- Local-first — The MCP server runs on your machine, not in the cloud
Getting Started
- Setup Guide — Generate a token and configure your IDE
- Tools Reference — Detailed documentation for each MCP tool
- IDE Integration — Specific instructions for Claude Code, Cursor, Windsurf
Example: How It Works in Practice
You’re working in Claude Code on your authentication module. You ask:
“What’s the auth pattern for this project?”
Claude Code calls chorum_query_memory with your question. ChorumAI returns:
{
"items": [
{
"type": "decision",
"content": "JWT with refresh tokens, httpOnly cookies for session management",
"relevanceScore": 0.92
},
{
"type": "invariant",
"content": "Never store tokens in localStorage—use httpOnly cookies only",
"relevanceScore": 0.87
}
]
}Claude Code now has the context to give you a response that matches your project’s established patterns.
“Use whatever IDE you want. ChorumAI is where your AI remembers you.”