Skip to content

Integrations

Proxy connects to external tools and services through a tri-modal integration system — three transport types unified under one interface.

Transport Types

1. MCP Servers

Persistent connections using the Model Context Protocol. Proxy discovers tools automatically via the MCP protocol handshake.

Proxy ←→ MCP Server (stdio or HTTP/SSE)
         ↓
    Tool discovery
    Tool execution
    Resource access

Examples: GitHub, Slack, Proxy's own MCP server, custom MCP servers.

2. CLI Tools

Subprocess invocation with structured JSON output. Proxy calls the tool as a child process and parses the result.

Examples: gh (GitHub CLI), gws (Google Workspace CLI), aws.

3. Direct HTTP APIs

REST calls with stored credentials. Proxy makes HTTP requests directly to external APIs.

Examples: OpenRouter, Stripe, custom webhooks.

Managing Integrations

Integrations are managed through the Proxy CLI or the Attic (settings) UI:

# List connected integrations
proxy integration list --connected

# Add a new integration
proxy integration add my-server

# Connect via stdio
proxy integration connect stdio my-server '{"command":"node","args":["server.js","--stdio"]}'

# Connect via HTTP
proxy integration connect http my-server '{"url":"http://localhost:3000/mcp"}'

# See available tools
proxy integration tools my-server

# Import from Claude Desktop config
proxy integration import

Harnesses

Harnesses are a special class of integration for agent runtimes — external programs that can participate in Party sessions as full agents.

Harness Transport What it is
Claude Code Process (stdin/stdout JSONL) Anthropic's CLI agent
Codex WebSocket OpenAI's coding agent
OpenClaw WebSocket Open-source agent runtime

Harnesses differ from regular integrations because they're bidirectional — Proxy sends prompts to the harness, and the harness sends tool calls and responses back.

# List available harnesses
proxy harness list

# Connect a harness
proxy harness connect claude-code

# Configure
proxy harness set-path claude-code /usr/local/bin/claude

# Check status
proxy harness status

Loadouts

A loadout is a saved configuration of integrations and skills for an agent. Each agent can have multiple loadouts that define what tools and MCP servers are available to them.

# List loadouts
proxy loadout list

# Equip an integration to a loadout
proxy loadout integration equip <loadout_id> <integration_id>

# Equip a skill
proxy loadout skill equip <loadout_id> <skill_id>

Proxy MCP Server

Proxy itself runs an MCP server on localhost:51711 when the app is open. This lets external tools interact with Proxy:

  • Read and write the Life Map
  • Manage agents and teams
  • Send messages to Party sessions
  • Take screenshots
  • Navigate the UI

Any MCP client can connect. This is how Claude Code and other tools integrate with Proxy for AI-assisted development.