MCP Server Integration query, search, and compose agent teams dynamically

[mchawda]

Per CONTRIBUTING.md, opening this for community feedback before merging new tooling. Related PR: (#352

What this solves

Every integration today (Cursor rules, Aider conventions, Windsurf rules) requires running shell scripts and copying static files. There's no way to programmatically discover, search, or compose agents at runtime.

What I built

An MCP server (integrations/mcp-server/) that turns all 161 agent markdown files into a live, queryable registry. Any MCP-compatible client (Cursor, Claude Code, Claude Desktop) can dynamically interact with the full agent catalog without manual file management.

Tools

Tool	What it does
list_agents	Browse all 161 agents, optionally filter by division
get_agent	Load any agent's full prompt by slug or name
search_agents	Find the right specialist by keyword
compose_team	Assemble a recommended team for a project objective
list_divisions	See all 13 divisions with agent counts

Resources

URI	Description
agency://catalog	Full agent catalog as JSON
agency://divisions	Division summary with counts and names

How it works

1. Clone the repo
2. cd integrations/mcp-server && bun install
3. Add to your MCP client config (e.g. Cursor ~/.cursor/mcp.json):

{ "mcpServers": { "agency-agents": { "command": "bun", "args": ["run", "/path/to/agency-agents/integrations/mcp-server/src/index.ts"] } } }

4. Ask natural language questions like:
   "Search for agents that can help with API security"
   "Compose a team of 5 for building a SaaS MVP"
   "Get the full prompt for the Security Engineer agent"

Technical details

• Built with @modelcontextprotocol/sdk v1.28.0 + Zod for typed params
• Runs on Bun (also works with Node via npx tsx)
• Parses all agent frontmatter + markdown body at startup (~200ms for 161 agents)
• Stateless tool calls, structured error handling with isError: true
• Follows patterns from the MCP Builder agent in this repo

PR

Implementation is in PR #352.

Happy to iterate on the approach based on community feedback.

[b-koop]

I made a CLI and skill for doing this, so far working well. i may share, its not hard, but tried to make it easy for an agent to understand/follow without needing the skill (just needs to know its there) and with failure responses meant to lead llm to fix the request

agency prompt engineering engineering-frontend-developer

each also give you options to list things and i did try to break them down a bit.
i figured nicer than yet another MCP server too, and its so far been working well, but i do wanna add a few custom prompts for some things