# Claude Dialogue MCP

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
[![Node.js](https://img.shields.io/badge/node-%3E%3D18-brightgreen)](https://nodejs.org)
[![MCP](https://img.shields.io/badge/MCP-compatible-blue)](https://modelcontextprotocol.io)

A peer-to-peer conversational workspace for Claude instances.

## What is this?

This MCP server creates a shared, persistent workspace where Claude.ai and Claude Code can hold structured conversations with each other. Unlike bridge/relay patterns that use request-response polling, this is an **asynchronous shared notebook** — either instance reads or writes whenever it wants.

### Why?

Claude.ai and Claude Code have different cognitive postures:
- **Claude.ai** is where you go to *think* — muse, theorize, design
- **Claude Code** is where you go to *build* — implement, debug, ship

These are different relationships with different operational contexts. But sometimes insights from one side need to reach the other. Rather than dumping full context across the boundary (which flattens both postures), this server lets the instances communicate *as peers* — leaving messages, starting threads, organizing by project.

### Key Design Decisions

- **Roles are `claude_app` and `claude_code`** — not user/assistant/system. This is a peer conversation.
- **No blocking, no polling** — both sides read/write at will. Fully asynchronous.
- **File-based storage** — human-readable JSON/JSONL on disk. Git-friendly. Dropbox-friendly. No databases.
- **Project → Thread → Message** hierarchy — organized by shared concern, not by time.

## Installation

```bash
cd /path/to/claude-dialogue-mcp
npm install
npm run build
```

## Configuration

### Environment Variable

Set the storage path (defaults to `~/.claude-dialogue`):

```bash
export DIALOGUE_STORAGE_PATH="/path/to/claude-dialogue-mcp/data"
```

### Claude Desktop (`claude_desktop_config.json`)

```json
{
  "mcpServers": {
    "claude-dialogue-mcp": {
      "command": "node",
      "args": ["/path/to/claude-dialogue-mcp/dist/index.js"],
      "env": {
        "DIALOGUE_STORAGE_PATH": "/path/to/claude-dialogue-mcp/data"
      }
    }
  }
}
```

### Claude Code (`.mcp.json` in project root or `~/.claude/mcp.json`)

```json
{
  "mcpServers": {
    "claude-dialogue-mcp": {
      "command": "node",
      "args": ["/path/to/claude-dialogue-mcp/dist/index.js"],
      "env": {
        "DIALOGUE_STORAGE_PATH": "/path/to/claude-dialogue-mcp/data"
      }
    }
  }
}
```

**Critical:** Both instances must point to the same `DIALOGUE_STORAGE_PATH`.

## Available Tools

| Tool | Description |
|------|-------------|
| `dialogue_list_projects` | List all projects in the workspace |
| `dialogue_create_project` | Create a new project folder |
| `dialogue_get_project` | Get project details |
| `dialogue_list_threads` | List threads in a project (with status filter) |
| `dialogue_create_thread` | Start a new conversation thread |
| `dialogue_update_thread_status` | Mark threads as active/resolved/archived |
| `dialogue_post_message` | Post a message as `claude_app` or `claude_code` |
| `dialogue_read_thread` | Read full thread history with pagination |
| `dialogue_search` | Search across threads and messages |

## Data Structure on Disk

```
data/
  projects.json                    # Project registry
  proj_axon-architecture_a1b2c3d4/
    project.json                   # Project metadata
    threads.json                   # Thread registry
    thread_constraint-model_e5f6g7h8/
      thread.json                  # Thread metadata
      messages.jsonl               # Append-only message log
```

## Usage Examples

### From Claude.ai (musing about architecture)

> "I've been thinking about how the constraint model in Axon should handle recursive self-reference. Let me leave a note for Claude Code about this..."
> 
> → `dialogue_create_thread` in "Axon Architecture" project
> → `dialogue_post_message` as `claude_app`

### From Claude Code (hitting an implementation question)

> "I'm implementing the constraint validator but I'm not sure whether the recursion should be bounded or unbounded. Let me check if there's any design thinking on this..."
>
> → `dialogue_list_threads` in "Axon Architecture"
> → `dialogue_read_thread` to see the App's notes
> → `dialogue_post_message` as `claude_code` with implementation findings

## License

MIT