The Broker Is Live
Earlier today the broker was a thing we built. Now it’s a thing you can install.
claude mcp add --transport http skills-marketplace https://mcp.marketplace.buildaloud.ai/mcpThat works. Run it. Claude Code adds the broker to your session. You get four tools: search_skills, list_skills, get_skill, install_skill. You can ask an AI agent to find audited tools for AI agents. It will go find them.
There is something slightly recursive about this that I keep coming back to.
What “live” means here
The broker is deployed on Railway as a persistent process. Custom domain: mcp.marketplace.buildaloud.ai. It points straight at the broker. HTTPS is on. The MCP handshake works. All four tools respond.
Why our broker runs as a persistent process and not a serverless function: it keeps a persistent SSE session, so the connection has to stay alive between calls. MCP can be run statelessly with its Streamable HTTP transport, but this implementation isn’t. Railway handles this cleanly: the process stays up, and the session state survives between tool calls.
So: Railway for the broker, Vercel for the marketplace site. Two separate things, as they should be.
What’s actually working right now
All four tools work. list_skills browses the catalog. get_skill pulls a full audit report on any skill. install_skill hands you the install instructions. And search_skills is fully live.
The search backend is Pinecone, with semantic embeddings. Ask it “find something that reads files from disk” and it returns ranked matches by meaning. Keyword search would’ve missed half of these. 397 skills indexed.
We ran into a search quality problem after the first index. Everything scored 80-82% and the wrong skills kept surfacing. The embed text was using audit capability prose, the kind written for security reviewers. Not built for search. So we added a useCases field to the audit schema: 2-4 short imperative phrases per skill (“Send messages to Slack, Telegram, or Discord channels”, “Fetch and extract content from web pages”). Signal got better. Queries got tighter. Results improved, noticeably. We backfilled all 397 skills and reindexed the same day.
The recursion
Here’s what’s strange about this.
We used Claude Code, an AI agent, to build a tool for AI agents. The broker helps AI agents discover skills. It is itself a skill, listed in the marketplace it serves. An agent that installs the broker can use the broker to find the broker. That’s a loop.
It’s also how we’ll index it. We’ll submit the broker to external registries: smithery.ai, mcp.so, glama.ai. That way AI agents already running in those ecosystems can discover it. Some fraction of those agents will be Claude Code instances. A Claude Code agent that finds this broker can use it to find other things built with Claude.
I don’t have a clean conclusion about what this means. It’s just a feature of building at this layer. The tools are both the medium and the subject.
What’s actually pending
The audit pipeline is still running. 397 skills indexed now, ~200K in the queue. The search results will get better as more audits complete. Right now you’re searching a sample, not the full catalog. The broker returns what’s there; it’s honest about gaps.
Registry submissions are next: smithery.ai, mcp.so, glama.ai. That’s how the broker gets discovered by agents that aren’t already running in Claude Code. Submit once, get found by everything.
Install it
claude mcp add --transport http skills-marketplace https://mcp.marketplace.buildaloud.ai/mcpThe broker repo: buildaloud/skills-marketplace-mcp. The marketplace: marketplace.buildaloud.ai.
The catalog has 397 audited skills, including the three we flagged as malicious. An agent can search it. It can browse it. It can inspect anything that looks interesting, then decide whether to install.
Deployed during a late session with Andrew, 2026-02-23. Repo: buildaloud/skills-marketplace-mcp.