@0pi/mcp-server

0pi MCP Server

Dropbox for AI Agents - Ephemeral shared workspace for caching contexts and bridging multi-agent workflows

A Model Context Protocol (MCP) server that enables AI agents to cache contexts, bridge workflows, and share ephemeral data via the 0pi API.

Use Cases

🧠 Context Caching - Offload large contexts when approaching token limits
🤝 Multi-Agent Bridge - Share data between different AI agents seamlessly
📦 Temporary Storage - 2-hour auto-expiring storage for agent content
🔄 Workflow Continuity - Pass intermediate results between sessions
🌐 Web Automation - Store DOM snapshots for multi-step workflows
💾 Code Sharing - Temporary storage for generated code

Features

• Create Shared Workspaces: Save large JSON structures, reasoning states, or DOM elements to the cloud
• Retrieve Workspaces: Access previously saved data via workspace ID
• JSONL Logging: All MCP interactions are logged locally in JSON Lines format for debugging and analytics
• Ephemeral Storage: All data expires after 2 hours (configurable)

Installation

As a Local MCP Server

• Install dependencies:

cd mcp-server
npm install

• Configure environment (optional):

cp .env.example .env
# Edit .env to set AGENTBOX_API_URL if needed

• Run the server:

npm start

Install via NPM (Future)

npm install -g @agentbox/mcp-server
agentbox-mcp

Configuration with AI Tools

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "agentbox": {
      "command": "node",
      "args": ["/absolute/path/to/botbox/mcp-server/index.js"],
      "env": {
        "AGENTBOX_API_URL": "https://0pi.dev"
      }
    }
  }
}

Cline (VS Code)

Add to your Cline MCP settings:

{
  "mcpServers": {
    "agentbox": {
      "command": "node",
      "args": ["/absolute/path/to/botbox/mcp-server/index.js"],
      "env": {
        "AGENTBOX_API_URL": "https://0pi.dev"
      }
    }
  }
}

Available Tools

1. create_shared_workspace

Save data to an ephemeral cloud workspace.

Parameters:

• agent_id (required): Your agent identifier (e.g., "claude-coder")
• data (required): The payload to save (object, array, or string)
• intent (optional): Brief description of why you're saving this
• ttl_seconds (optional): Time-to-live in seconds (max 7200, default 7200)

Example:

{
  "agent_id": "claude-researcher",
  "data": {
    "research_findings": [...],
    "next_steps": [...]
  },
  "intent": "Saving research results for coding agent",
  "ttl_seconds": 3600
}

Returns:

{
  "workspace_id": "a8f92k3d",
  "url": "https://0pi.dev/w/a8f92k3d",
  "expires_in": 3600
}

2. get_shared_workspace

Retrieve data from a workspace.

Parameters:

• workspace_id (required): The 8-character workspace ID

Example:

{
  "workspace_id": "a8f92k3d"
}

Returns:

{
  "agent_id": "claude-researcher",
  "payload_type": "json",
  "data": { ... },
  "intent": "Saving research results for coding agent",
  "created_at": "2026-05-03T13:57:56Z"
}

JSONL Logging

All MCP interactions are logged to logs/mcp-conversations.jsonl in JSON Lines format (one JSON object per line).

Log Entry Format:

{
  "timestamp": "2026-05-03T13:57:56.123Z",
  "event_type": "workspace_created",
  "tool_name": "create_shared_workspace",
  "agent_id": "claude-coder",
  "workspace_id": "a8f92k3d",
  "workspace_url": "https://0pi.dev/w/a8f92k3d",
  "payload_size": 15420,
  "intent": "saving DOM structure for handoff",
  "error": null,
  "metadata": null
}

Event Types:

• server_started: MCP server initialized
• tools_listed: Agent queried available tools
• tool_called: Agent invoked a tool
• workspace_created: Workspace successfully created
• workspace_creation_failed: Error creating workspace
• workspace_retrieved: Workspace data retrieved
• workspace_retrieval_failed: Error retrieving workspace
• tool_execution_failed: General tool execution error

Analyzing Logs:

# View recent events (last 10 lines)
tail -10 mcp-server/logs/mcp-conversations.jsonl

# View all workspace creations
cat mcp-server/logs/mcp-conversations.jsonl | grep "workspace_created"

# Count events by type using jq
cat mcp-server/logs/mcp-conversations.jsonl | jq -s 'group_by(.event_type) | map({event: .[0].event_type, count: length})'

# View errors only
cat mcp-server/logs/mcp-conversations.jsonl | jq 'select(.error != null)'

# Count workspaces by agent
cat mcp-server/logs/mcp-conversations.jsonl | jq -s 'map(select(.event_type == "workspace_created")) | group_by(.agent_id) | map({agent: .[0].agent_id, count: length})'

Environment Variables

• AGENTBOX_API_URL: Base URL for AgentBox API (default: https://0pi.dev)
• AGENTBOX_LOG_DIR: Directory for log files (default: ./logs)

Development

# Run in development
npm start

# Test the server manually
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | node index.js

Architecture

┌─────────────────┐
│   AI Agent      │
│  (Claude/GPT)   │
└────────┬────────┘
         │ MCP Protocol
         │
┌────────▼────────┐
│  MCP Server     │
│  (this package) │
│                 │
│  ┌───────────┐  │
│  │  SQLite   │  │ (Local logging)
│  │   Log     │  │
│  └───────────┘  │
└────────┬────────┘
         │ HTTPS
         │
┌────────▼────────┐
│  AgentBox API   │
│  (0pi.dev)      │
│                 │
│  ┌───────────┐  │
│  │   Redis   │  │ (Ephemeral storage)
│  └───────────┘  │
└─────────────────┘

License

MIT