Agents use MCP tools — filesystem, browser, databases, any MCP server
Council agents can call any MCP server during deliberation. File system access, browser automation, database queries, Slack messaging — any tool available via MCP becomes available to your agents. The same protocol Claude and other AI tools use.
MCP tools appear as callable functions. Agent writes TOOL_REQUEST: mcp_server.tool_name params → Council routes to the MCP server → result injected into next round.
Council deliberates a technical architecture decision. Agents use MCP to access real project data.
MCP: filesystem.read_file("src/auth/middleware.py")Read the actual implementation — found 3 security patterns that constrain the redesign.
MCP: postgres.query("SELECT avg(latency) FROM requests WHERE path LIKE '/auth%'")Auth endpoints average 340ms — already slow. Proposed change would add another layer.
MCP: confluence.search("auth migration plan")Found existing migration plan from Q1 — 60% overlaps with this proposal. Don't duplicate.
RECOMMENDATION: Adapt the Q1 migration plan instead of starting fresh. Address the 340ms latency as part of the migration, not as a separate effort. The 3 security patterns in middleware.py must be preserved.
In ecosystem settings, add MCP server configs: command, args, env vars.
{"name": "project_fs", "type": "mcp", "transport": "stdio", "mcp_command": "npx", "mcp_args": ["-y", "@anthropic-ai/mcp-server-filesystem", "/path/to/project"]}Council calls list_tools() on each MCP server at startup. Tools appear as agent capabilities.
Each agent's ## Tools section whitelists which MCP tools they can call.
Anthropic's official servers, community servers, or your own custom MCP implementations.