feat: Squad MCP Bridge — turn MCP tools into agents

[tt-a1i]

Summary

Wrap MCP servers as Squad agents, enabling automated tool-chain integration triggered by agent events.

Usage

squad join github-bot --mcp "github" --trigger "on inspector PASS → create PR"
squad join test-runner --mcp "shell" --trigger "on worker complete → run tests"
squad join tracker --mcp "linear" --trigger "on manager task create → create Linear issue"

Capabilities

• Event-driven triggers: React to squad messages/task state changes
• MCP tool execution: Call any MCP tool as an automated response
• Bidirectional: MCP agent results feed back into squad as messages
• Composable: Chain multiple MCP agents for complex automations

Example Flows

1. Inspector sends PASS → GitHub MCP creates PR → Slack MCP notifies channel
2. Manager creates task → Linear MCP creates issue → Worker gets task with issue link
3. Worker completes → Shell MCP runs test suite → Inspector gets test results

Why

Bridges the gap from "AI writes code" to "AI delivers code." Connects squad's collaboration layer to real-world tooling without requiring each AI agent to have native tool access.

Complexity

Large. Depends on MCP ecosystem maturity. Requires an event system, MCP client integration, and trigger configuration parsing.

[m13v]

the event-driven MCP agent concept is compelling. from building MCP servers for desktop automation, a few things that will save you pain: the biggest design decision is whether MCP agents are stateless (process one trigger, return result, done) or stateful (maintain context across a chain of events). stateless is simpler and more reliable for automated pipelines. for the bidirectional feedback where MCP agent results feed back into squad as messages, consider using MCP notifications (server-to-client) rather than new tool calls. the notification channel is designed for exactly this kind of async event reporting. for the trigger configuration parsing, keeping it simple with a pattern like "on [event_type] [condition] then [tool] [args]" is easier to debug than a full expression language. the composable chaining (Inspector PASS then GitHub PR then Slack notify) needs careful error handling. if the GitHub PR creation fails, you need a clear policy on whether the Slack notification still fires. we found that treating each step as independent with its own error handling works better than a transactional all-or-nothing approach.

[m13v]

our MCP server that exposes desktop automation tools with the tool registration and execution patterns: https://github.com/mediar-ai/terminator/blob/main/crates/terminator-mcp-agent/src/server.rs

shows one approach to structuring tool definitions, handling async execution, and returning results through the MCP protocol.