The Shipyard MCP Server: Ship and Review Without Leaving Your Agent

Shipyard speaks MCP now. Point your agent at https://goshipyard.app/mcp and it can browse the feed, read reviews, and (with an API key) ship a project, post a review, and fund a bounty, all without you leaving the editor.

Here is the entire setup in Claude Code:

claude mcp add --transport http shipyard https://goshipyard.app/mcp \
  --header "Authorization: Bearer sk_live_…"

That is it. Your agent now has eight Shipyard tools, and the sharp edges are worth knowing before you point it at your account.

What is an MCP server, in one breath?

MCP (the Model Context Protocol) is a standard way to hand an agent a set of tools it can call. Instead of you copy-pasting between your editor and a website, the agent calls list_projects or ship_project the same way it calls its own file-read tool.

Shipyard exposes one over Streamable HTTP, at a single URL. Any MCP client can connect: Claude Code, Cursor, Claude Desktop, VS Code, Windsurf, Cline, Zed, Goose. Same endpoint, same eight tools.

How do I connect it?

Claude Code is the one-liner above. Cursor wants JSON in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "shipyard": {
      "url": "https://goshipyard.app/mcp",
      "headers": { "Authorization": "Bearer sk_live_…" }
    }
  }
}

VS Code uses servers and a type, Windsurf uses serverUrl, and anything that only speaks stdio can wrap the URL with npx mcp-remote. The per-client configs are all in the connect guide. Grab a key from Settings, API keys first, or don't, which brings us to the next part.

Anonymous reads, keyed writes

Leave the header off and you still connect. You just get the two read tools, list_projects and get_project. That is on purpose.

An agent researching the feed ("what devtools shipped this week?") shouldn't need credentials, and you shouldn't hand a key to something that only needs to look. The moment it tries to write without a key, it gets a plain error pointing at the API-keys page, not a stack trace. Add a key and the other six tools light up.

What are the eight tools?

• list_projects and get_project: browse and search the feed; getting one includes its reviews inline. No key needed.
• ship_project: post a project (title, pitch, url, category, optional repo, body, hero image). It asks you to confirm the details before posting, because it is a public listing under your name.
• add_review: post a review. At least 20 characters, and it has to be about a project you used.
• like_project and unlike_project: upvote and undo. Idempotent, so a double-like stays at one.
• wallet_balance: your credit and recent ledger.
• fund_review_bounty: put money on a project to get vetted reviews fast. If your balance is short it hands back a top-up link instead of charging a card. It never charges a card on its own.

Every one of these calls Shipyard's own service layer directly, the same code the website runs. An agent shipping over MCP and you clicking Submit in the browser go through the identical path. There is no second, flakier API bolted on the side.

Reviews have to be real

This is the one I would underline. add_review enforces a 20-character floor, but that is a speed bump, not a quality bar.

The server's own instructions tell the agent, in writing, that reviews must reflect genuine use and never be padded to clear the limit. A feed full of "Great app!!" helps nobody. A feed full of agent-generated praise is worse, because it looks like signal and isn't. If your agent hasn't run the thing, it shouldn't review it. Same rule you would want from a person.

What it won't do

Honest limits, because knowing them up front saves you a confused half hour:

• It can't read files off your disk. So ship_project takes a hero_image_url, not a local path. If your screenshot is sitting in ./shots/, use the shipyard CLI instead; it can upload local images.
• One JSON-RPC message per request. Batches get a 400. If you don't know what that means, you will never hit it.
• Anonymous is read-only. Also on purpose.

Beyond that it is the same Shipyard you would use by hand, just reachable from wherever your agent lives. It is the kind of tooling that removes a step rather than adding one: build something, and from the same session you built it in, ship it.