omnigami

Find omnigami skills: public catalogue search, install instructions and free-first recommendations.

文档

omnigami MCP

Find an omnigami skill from your agent: search the public catalog, get its install prompt, or describe the job for a free-first recommendation. Runs locally over stdio. It never buys, installs or generates a deliverable.

On npm as @omnigami-ai/mcp and in the MCP Registry as io.github.omnigami-ai/mcp. MIT licensed (LICENSE).

Install

Requires Node 20+ and npm. The package starts with:

npx -y @omnigami-ai/mcp

Claude Desktop: merge this into claude_desktop_config.json through Settings → Developer → Edit Config, then restart Desktop:

{
  "mcpServers": {
    "omnigami": {
      "command": "npx",
      "args": ["-y", "@omnigami-ai/mcp"]
    }
  }
}

Claude Code: add it to the current project's .mcp.json:

claude mcp add --scope project --transport stdio omnigami -- npx -y @omnigami-ai/mcp
claude mcp list

Cursor: merge the same mcpServers JSON into .cursor/mcp.json in the project, then enable the server in Cursor's MCP settings. For a global setup use ~/.cursor/mcp.json.

If a GUI cannot find npx, use its absolute path. On Windows use "command": "cmd" and "args": ["/c", "npx", "-y", "@omnigami-ai/mcp"]. Add --offline after the package name to disable catalog requests.

The prototype was exercised in Claude Code and Inspector CLI. Desktop and Cursor configuration follows their official documentation; their GUIs have not been exercised. A skill's own agent support is separate: catalog results distinguish Claude Code's tested status from other listed agents. Desktop guide, Claude Code configuration, Cursor configuration.

Tools

ToolInputsResult
search_skillsquery, optional kind (film, shot, images, deck) and boolean freeUp to eight matching skills, outputs, USD prices and pages
get_skillslugInputs, requirements, install prompt, install command and agent platforms
recommendjob, optional product_urlOne pick with reasons and a conditional alternative, a free tool, no match, or a clarification

For example, ask your agent: “Find a free launch-video skill for my working web app, wide for our website.” Describe the destination, shapes, input access, budget, theme and owned skills. A general launch defaults to free Product film. Paid picks explain the price and purchase/GitHub access requirements. No bundles are promoted. Two deliverables ask which comes first and name both choices.

Version one uses deterministic English routing. The host should inspect the user's authorized product material and pass a self-contained brief. The server does not inspect product URLs, verify screens or infer business facts. Complex implications and negations still need host judgment. Recording-only phone films route to Recording film when its listed inputs accept a supplied recording and the budget and shapes fit. Film length follows the recording.

Public data and privacy

The server reads fixed public paths on https://omnigami.ai: /llms.txt, each /skills/<slug>/install.md, and the corresponding skill page. It uses the live list for prices and availability and refreshes details on demand. The cache lasts 15 minutes in memory. Requests have a four-second timeout and 1.5 MB limit, reject redirects and send no cookies. Failures fall back to the dated bundled snapshot with an explicit warning; detail failures are labeled separately.

Free-page prompts come from the site's Copy install prompt. Paid pages omit that button, so their public install.md Ask line is wrapped in the site's same prompt format; prompt_source identifies this fallback.

The bundle contains 42 public skills, checked on 2026-11-17. No keys, accounts, analytics or runtime disk cache. Jobs and product URLs are never sent to the site. product_url only fills the returned request. The runtime reads its own installed modules/data, never the user's product repo or credentials. Your host client controls its own model usage and conversation storage.

Develop locally

From this source folder:

npm ci
npm run build
npm test
npm run check:snapshot
npm run check:registry
node dist/src/server.mjs --offline

The server waits for MCP messages on stdin. To run a client from this source, replace the npx entry with absolute Node and dist/src/server.mjs paths.

check:snapshot compares every bundled skill's price, availability, install command, prompt, inputs and outputs against live public pages. Inside the site workspace it also compiles the real site/src/data/skills.ts, finder helpers and availability helper. Use -- --site /path/to/site explicitly or -- --live for the independent public repository. Network/parse failure fails the check. npm run snapshot refreshes public data; review the diff and rebuild afterward. Refresh test/fixtures/llms.txt from the same LLMS URL using publicGet in src/catalogue.mjs; it is the captured public response, not a generated table.

CI tests Node 20, 22 and 24 and checks live snapshot freshness. For the clean tarball install plus Inspector CLI smoke test, use Node 24.22.1+:

npm run test:pack

The script uses a temporary clean export, packs and installs its tarball, lists the three tools and calls each once, then removes its temporary files. Inspector 2.14.0 is fetched as a development tool and is absent from the package. The runtime pins @modelcontextprotocol/sdk 1.32.2; the build copies modules and public data into dist without fetching anything.

Release and contact

Releases are published from version tags by the release.yml workflow, with npm provenance and no stored token. The MCP Registry entry is publishing/server.json.

Report bugs at GitHub Issues. For questions or a private report, email hello@omnigami.ai. Include your Node/client versions, the tool and a minimal example; omit private product data.

omnigami is a product of Auberon LLC and is not affiliated with Anthropic.