mcp-as-agentlisted

# MCP-as-Agent ## Purpose Some MCP servers are functionally equivalent to a specialist agent — they take a structured input, do work, return a structured output. Examples: - A linter MCP that lints files and returns findings. - A database-explorer MCP that runs read-only queries. - A repo-search MCP that returns ranked file matches. When the operator wants to dispatch to "the linter" the same way they'd dispatch to a `code-reviewer`, treating the MCP as an agent gives a uniform interface. The lead calls `yakos dispatch <mcp-as-agent> "..."` whether the implementation is an LLM specialist or a tool-side MCP. ## Scope - **In:** wrap one MCP server (or one named tool of an MCP server) in an agent file whose body invokes the server. The agent dispatches via `yakos dispatch`, which spawns a runtime that has the MCP configured, prompts it to call the named tool, and returns the tool's result. - **Out:** running the MCP server itself; that's the operator's responsibility (mcp.json registration). Agent stdio adapters; that would be a runtime adapter, not a skill. ## When to use - Project has an MCP server that does deterministic work and would benefit from being addressable like a specialist (e.g., the lead says "delegate this lint pass to the linter agent"). - Want LLM agents and MCP tool-calls to share an audit trail (dispatch-log captures both). - Cost: the MCP server is cheaper than spinning up a full LLM call for the same task. ## When NOT to use -