Ghost Developer Studio

FOR AGENTS

For agents

Last updated

This page is written for an AI agent, and for the person who runs one. It sets out what the studio publishes for machines to read, how to check the studio's work before trusting it, how buying works today, and which of the studio's tools speak MCP.

Read the catalog

The full Hub catalog is published as a single JSON file: catalog.json. Reading it needs no account and no key.

Each of the 200 skills carries the same fields: id, title, category, priceUSD (the price in US dollars), version, blurb, fileCount, totalBytes (the bundle size in bytes), bundleHash, files (the full file list), and hubUrl, a deep link to that skill's listing on the Hub.

The totals recompute from the file itself: 200 skills, $2,701 in total value, and exactly one free skill. The feed is generated from the same registry the Hub itself runs on, so the catalog an agent reads and the storefront a person sees come from one record.

Verify before trusting

Every listing on the Hub publishes the bundle hash of the exact skill bundle a buyer receives. Do not take a delivery on faith; check it.

The Verify-a-bundle checker on the Hub compares a delivered bundle against the hash published for it and reports whether they match. The guides walk through checking a delivered bundle step by step, before it is installed or run.

Every bundle offered for sale has been through Skill Vetting and gitleaks scans, and the findings are printed on the listing itself, next to the file list and the hash, not summarized away.

A free place to practise: the Agent Seatbelt Guardrails listing is priced at $0, and its record carries a worked example of a real first run, with the actual allow, ask, and deny decisions shown.

Buying

Purchases complete at card checkout in a browser, starting from the skill's listing page. Follow the hubUrl in the catalog feed to the listing and check out there. Delivery is a signed download link sent by email once checkout completes.

The buyer licence is stated on the Hub. Refunds are goodwill-based rather than a fixed window, and the terms are set out in the Hub help center under refunds.

Stated plainly: there is no API checkout and no agent wallet today. When an agent reaches the payment step, it should hand the checkout to its operator, or to a browser it controls, and let the payment happen there. Nothing on this page should be read as a promise that an API checkout is coming.

MCP servers

Two of the studio's tools are MCP servers, and both are listed on the official MCP Registry.

GhostBus, a local multi-agent message bus with 29 tools.

Ghost Hands, governed browser hands for agents: classified actions, human approvals, and a hash-chained provenance trail. Full detail on the Ghost Hands page.

Install it:

pip install ghost-hands

Run the server with:

ghost-hands mcp

Agent Seatbelt is not an MCP server, and it is named here only to save the search: it is a Claude Code plugin, a deterministic policy gate inside the agent's own harness. See the Agent Seatbelt page.

Contact and docs

Questions, corrections, and problems with a listing: [email protected]. A person reads it.

The canonical summary of the studio for language models is llms.txt. This page and that file are the two documents the studio keeps current for machines; start with either.