MCP server for any AI client

Building a correct machine profile means knowing each tool's exact name and how every operating system installs it. People already work with an AI they like, but a model left to guess invents package names and flags, and ripeat changes real machines. The requirement was an interface an agent can rely on for lookups, validation and risk, without the authority to act unchecked.

I designed ripeat's own MCP server around single-purpose tools: read the machine, resolve names, look a tool up, validate a list, plan it, rate its risk and save a named profile. A build-profile prompt starts the flow, and execution stays out of the default surface. Two run tools exist only when the server is started with --allow-run; each call is gated by an approval prompt the model cannot answer, and a high-risk plan, a blocked tool or a step that needs administrator rights is refused regardless of the conversation. Recipes for unknown tools are generated by ripeat under fixed rules and approved by a person, never written by the model. ripeat mcp setup connects Claude Code, Claude Desktop or Codex in one command, and other MCP clients register the same server.

Designed and built an MCP server with 26 purpose-built tools that lets Claude, Codex or any MCP client assemble, validate and risk-rate a machine profile from a single request. Installation stays opt-in and gated by explicit human approval, and a high-risk plan is always refused.

• Execution is opt-in, so users who want agent-driven installs make an explicit choice.
• Approval depends on the client rendering a prompt; when it cannot, ripeat refuses and returns the command instead.

A profile that used to require looking up every tool by hand is now a single request to the client of choice, validated and risk-rated before anything runs, with the decision to install left to the person. The tool contracts are regression-tested against twelve scenarios.