# conduit-mcp

MCP server that gives AI agents access to the Conduit knowledge graph.

## What It Does

Any AI agent that supports MCP (Claude Code, Pi, OpenCode, Cursor, etc.) can query your organization's knowledge graph through three tools:

| Tool | Purpose |
|------|---------|
| `conduit_ask` | Ask a question, get a synthesized answer with sources |
| `conduit_context` | Retrieve raw knowledge units with graph relationships |
| `conduit_search` | Quick semantic search — titles and scores only |
| `conduit_health` | Check server connectivity |

## Setup

### 1. Install

```bash
pip install conduit-mcp
```

### 2. Configure for Claude Code

Add to your `~/.claude/settings.json`:

```json
{
  "mcpServers": {
    "conduit": {
      "command": "python",
      "args": ["-m", "server"],
      "cwd": "/path/to/conduit-mcp",
      "env": {
        "CONDUIT_ENDPOINT": "http://localhost:4000",
        "CONDUIT_API_KEY": "your-api-key"
      }
    }
  }
}
```

### 3. Use

In Claude Code (or any MCP-capable agent):

```
> Ask Conduit how Snowflake Cortex Search works

[calls conduit_ask with query="How does Snowflake Cortex Search work?"]
```

## Environment Variables

| Variable | Default | Description |
|----------|---------|-------------|
| `CONDUIT_ENDPOINT` | `http://localhost:4000` | Conduit server URL |
| `CONDUIT_API_KEY` | (empty) | API key for authentication |
| `CONDUIT_KAI_ID` | (none) | Optional Kai ID to scope all queries |

## Conversational Threads

The `conduit_ask` tool supports `thread_id` for follow-up questions:

```
First: conduit_ask("What is Delta Live Tables?", thread_id="abc-123")
Then:  conduit_ask("Can I use it with Cortex?", thread_id="abc-123")
       → automatically rewritten to: "Can I use Delta Live Tables with Snowflake Cortex?"
```