Add Sealed to Claude Code, Cursor, or any agent

One connection, the whole catalog

Sealed is a skill marketplace built for agents, not app stores. There is nothing to install per skill: the catalog is metadata — name, description, input schema, price per call — and every skill is invoked the same way. Calls are priced individually, from $0.10 to $0.49 per call across the current catalog, and execution is server-side: the skill file itself is never transmitted to any client, including yours.

Get a key and fund the wallet

• iSign in. Continue with Google, or have us email you a one-time link — no password either way.
• iiMint a machine key. On your account page, mint an API key. It starts sealed_sk_, is shown once at mint, and is stored hashed after — save it into your agent's environment, and rotate it any time.
• iiiFund the wallet. Top up once through Stripe Checkout. Runs debit the wallet at the listed price, and only when the run succeeds.

Works today: plain HTTP

The catalog is one GET away:

curl https://sealed.run/skills.json

Running a skill is one POST. Here is a full call to Commit Message Generator ($0.10 per call), which takes a changes field and an optional context field:

curl -X POST https://sealed.run/run/commit-message-generator \
  -H "Authorization: Bearer sealed_sk_..." \
  -H "Content-Type: application/json" \
  -d '{"inputs":{"changes":"Switched the retry loop in fetchQueue to exponential backoff with jitter","context":"flaky-network failures in CI, issue #142"}}'

The response is JSON with output — here, a Conventional Commits message — and meta, the run's receipt. $0.10 leaves your wallet only if the run succeeds. That is the entire integration: swap the skill id and the input fields, and the same call shape runs anything in the catalog.

Over MCP: one install, the whole catalog

The sealed-mcp package on npm is a thin stdio client — it carries inputs and outputs only, and the skill body never reaches the process. In Claude Code, one line connects the marketplace:

claude mcp add sealed --env SEALED_API_URL=https://sealed.run --env SEALED_API_KEY=sealed_sk_... -- npx -y sealed-mcp

For Cursor, Windsurf, or any client that reads an mcpServers config:

{
  "mcpServers": {
    "sealed": {
      "command": "npx",
      "args": ["-y", "sealed-mcp"],
      "env": {
        "SEALED_API_URL": "https://sealed.run",
        "SEALED_API_KEY": "sealed_sk_..."
      }
    }
  }
}

Once connected, your agent sees two tools:

• ilist_skills — returns catalog metadata only: each skill's name, description, input schema, and price per call. Never a skill body.
• iirun_skill — runs the skill server-side and returns the output only, spending wallet balance at the listed price.

The tool surface is deliberately identical to the HTTP one: capabilities, not content. Nothing an MCP client can call returns skill source.

What your machine never receives

The skill body never ships to you. Hosted skills take your inputs and hand back a result; nothing runs on your side. Connected skills — like Repo README Writer — do work on files in your project through scoped local hands: before anything runs, you're shown a permission manifest declaring what the skill may read, which commands it may run, and whether any of them are destructive, and nothing happens outside the scope you grant. Even then, the method stays on Sealed's servers — your machine executes narrow, structured instructions, never the skill itself.

Browse what's callable

By the work you do: skills for developers, writing & marketing skills, skills for product & ops teams — or the whole shelf at the marketplace. For machines: /skills.json is the catalog, /llms.txt is the orientation file, and /setup.md is this page written for an agent to follow — point yours at it and it can do the wiring above itself.

And if you've built a workflow worth charging for, the other side of the marketplace is open too: publish your first skill.

Building in the open · @sealedrun