Registry MCP server

Search and install items by asking your AI tool, with a plan you see first.

From the Documentation in Ronne AI Marketplace 0.3.0.

What it does

The registry MCP server, rmk-mcp, lets your AI tool (Claude Code, Codex, Cursor, or any tool that speaks MCP) do what rmk does, from the conversation. Ask it to find a skill for reviewing SQL, to explain what an item would change, or to update everything, and the assistant uses the server's tools.

It runs on your machine, started by the AI tool in your project folder, and reads this marketplace as you, with the token from rmk login. It uses rmk's own code, so it resolves, renders and writes exactly as rmk does, and rmk.lock and the state file stay the same whichever you use.

It can also send an item you wrote in your AI tool to this marketplace as a private draft, as rmk export does. Reviewing, submitting and releasing stay in this website.

Setting it up

Get it. npm install --global @ronneai/rmk @ronneai/mcp installs rmk and rmk-mcp (Node.js 22.12 or later). Running a copy of the repository instead? After pnpm build the server is packages/mcp/dist/bin.js: register it by its path with --command, below.
Log in with rmk login, if you haven't: the server uses that token.
Register it with each AI tool, from the project folder:

rmk mcp-setup                # for the tools this project uses
rmk mcp-setup --target all    # for every tool rmk knows
rmk mcp-setup --scope user    # in your home folder, for every project
rmk mcp-setup --command "node /path/to/packages/mcp/dist/bin.js"
rmk mcp-setup --remove

rmk mcp-setup adds the server to each tool's MCP settings as ronne-registry (.mcp.json for Claude Code, .codex/config.toml for Codex, .cursor/mcp.json for Cursor), and records it like any setting rmk writes: installs leave it alone, --remove takes exactly it away, and it never replaces an entry you made.

Then restart the tool, or reload its MCP servers. Claude Code asks once before it uses a project's MCP servers, and Codex reads .codex/config.toml only in a project you trust.

The tools

Tool	What it does	Writes or sends
search_items	Finds items by name, description or keyword; by type, scope or AI tool too.	Nothing
get_item	An item's tags and versions, and a version's dependencies, risk flags, the tools it works in and README.	Nothing
list_installed	What the lockfile holds, and which items you asked for.	Nothing
check_outdated	For each item you asked for: locked, newest its range allows, newest published.	Nothing
plan_install	Plans installing items, with @tag or @range if wanted.	Nothing
plan_update	Plans updating items, or all of them, within their ranges.	Nothing
plan_remove	Plans removing items, and what nothing else needs any more.	Nothing
apply_plan	Writes a plan made in the last 10 minutes, once.	The plan's files, the lockfile and the state
list_local_items	The skills, agents, commands, rules and MCP servers in Claude Code's, Codex's and Cursor's folders, with the tool each is for, and whose each is: yours, installed (edited or not), a registry copy, or written by rmk. Takes type and from.	Nothing
plan_export	Plans sending items of yours to this marketplace as drafts: every file, and what's left out. Takes type, from (the tool), an MCP server's description, and dependencies (include or omit) for the items of yours they use. An edited install, or an item whose name is published, is planned as a change proposal unless new is set.	Nothing
export_items	Uploads an export plan made in the last 10 minutes, once, as private drafts.	Sends the plan's files to this marketplace
check_drafts	Says which of your drafts Submit would take now, and what's in the way of the others. Takes items (names or ids) or all; your drafts they depend on are included first, unless dependencies is false.	Nothing
submit_drafts	Submits each of those drafts that's ready, and says why the others weren't.	Sends them to reviewers

Three tools act, and your AI tool asks you about them: apply_plan on your files, and export_items and submit_drafts on this marketplace. Every other tool is marked as read-only. list_installed, check_outdated and the plan tools take scope: user for your home folder, and the plan tools take targets, such as ["codex"], when the folder looks like several tools.

Plans

Plan. Asked to install, update or remove, the assistant first makes a plan: the versions, every file and setting it would write or remove, warnings about what a tool can't take, what each new item can do (its risk flags), and the environment variables its MCP servers need. Planning writes nothing to your project.
Apply. Once you've seen the plan, the assistant applies it with apply_plan; your tool asks you before it runs, unless you've allowed it to. It writes exactly that plan, then the lockfile and the state file, and says what it wrote.

A plan lasts 10 minutes, and is applied once.
If anything the plan touches changes in between, such as an install from the terminal or another AI tool, it's refused as stale, and the assistant plans again.
A plan with conflicts (a file or setting rmk didn't write, or that changed since it did) can't be applied from the AI tool: move them aside, or use rmk install --force at the terminal, on purpose.

Export plans work the same way: plan_export shows every file that would be uploaded and sends nothing, and export_items uploads exactly that plan, once, within 10 minutes. If you edit the skill in between, the plan is refused as stale. Install plans and export plans are kept apart: neither tool takes the other's.

What it can reach

It can do what your token can: read what's published and download it, create drafts in your name, and submit them when you ask. It never shows the token, and it can't sign in to this website.
It sends only items it found in your AI tools' folders and that are yours, never a folder you name, never the files export leaves out or an MCP server's credentials, and only when you approve export_items. What arrives is a draft only you see.
It writes only in the project folder (or your home folder, with scope: user), through the same checks as rmk, and never runs an item's code: hooks and scripts are written, not run.
Without a token, every tool that reaches the marketplace says to run rmk login. If the marketplace can't be reached, list_installed and list_local_items still work: they only read your files.