MCP-Anything
A tool that turns a codebase, an OpenAPI spec, a gRPC proto file, a GraphQL schema, or a plain-language brief describing a data source into a complete, working MCP server, without the developer hand-writing the tool definitions.
MCP-Anything works in two modes: a brief-driven build mode, where a language model reads a written brief describing a domain and its data source, groups related operations into ergonomic tools, writes agent-optimized tool descriptions, and generates the full server package including a SKILL.md agent guide and an evaluation harness; and a generate mode that scans an existing codebase, OpenAPI spec, gRPC proto or GraphQL schema directly with static detection and no LLM call at all. A Claude Code skill installs the underlying CLI automatically on first use, so a request such as turning an OpenAPI spec into an MCP server, or wrapping a Flask app as MCP tools, picks the right mode and runs it without a separate install step.
The generated package's key output, SKILL.md, is written for an AI agent rather than a person: it lists every tool with its parameters, operation modes, usage recipes, known gotchas and anti-patterns, which the project states an agent reads before making any calls to reduce hallucinated tool usage. Its tool-design rules include grouping three or more operations on the same resource into a single tool with an operation parameter, rather than exposing each one separately.
It installs from PyPI, requires an Anthropic API key for the brief-driven mode, is written in Python, and is released under the Apache-2.0 license.