Publishing a skill — LoopSkill Docs

LoopSkill is an open marketplace. Any developer can submit a skill for review. The canonical publish path is MCP-first — your agent calls recipes_publish_request from within a conversation. A CLI fallback is also available.

New in Phase C / recipes_2005

The recipes_publish_request MCP tool is live. It submits your SKILL.md directly from the agent conversation, runs the quality gate in-memory, opens a GitHub review issue, and returns a request_id and issue_url in one tool call. No dashboard, no upload form.

What gets published vs. what you keep

Published: your SKILL.md, optional references/ and templates/ directories, a skill.toml manifest.
Private: your env vars, your credentials, your agent config. None of those travel with the skill.
Revenue share: accepted skills earn a percentage of subscription revenue attributed to your skill's install and use events. See Creator program for current rates.

MCP-first path (recommended)

Step 1: Develop locally

Scaffold your skill directory using recipes_recipify or by hand:

SKILL.md          ← required: YAML frontmatter + instructions body
skill.toml        ← required: metadata (slug, version, tier, license)
references/       ← optional: reference docs, loaded on demand
templates/        ← optional: config snippets, scaffolds

The SKILL.md frontmatter must include at minimum: name, description, version (semver), tier (free | cook | operator), and license (SPDX).

Step 2: Call recipes_publish_request

From inside your agent conversation (e.g. Claude Code, Hermes):

// Tell your agent:
"Submit my skill at ./skills/my-skill/ to the LoopSkill marketplace"

// The agent calls:
recipes_publish_request({
  slug: "my-skill",
  content: "<SKILL.md contents>",
  version: "1.0.0",
  description: "What this skill does",
  tier: "cook",
  license: "MIT"
})

The tool runs synchronously and returns:

{
  "request_id": "uuid...",
  "slug": "my-skill",
  "status": "pending_review",
  "issue_url": "https://github.com/wisechef-ai/recipes-api/issues/NNN",
  "sha256": "abc123...",
  "warnings": []   // or [{severity: "low", message: "..."}]
}

If the quality gate finds a HIGH severity issue (e.g. curl | bash pattern, hardcoded $HOME path, credential leak), the tool returns {"error": "quality_gate_failed"} and no issue is opened — fix the finding and resubmit.

Step 3: GitHub review

The issue_url in the response links to a GitHub issue in the wisechef-ai/recipes-api repo. A WiseChef team member reviews:

Does the skill do what it says?
Is the install reproducible on a clean machine?
Does telemetry surface the expected success/failure events?
Is the quality gate fully green?

You can track the issue directly. The reviewer may leave comments asking for changes. Respond by updating your SKILL.md and calling recipes_publish_request again with the same slug — the system deduplicates and links the new submission to the existing issue.

Step 4: Signed and published

Once approved, a WiseChef team member closes the issue with the approved-publish label. An automation job:

Builds the final tarball from the approved submission
Signs with the WiseChef ed25519 publish key
Uploads to the CDN
Creates the skill row in the catalog database
Sends you a confirmation email with the published URL

CLI fallback

If you prefer a terminal workflow, the recipes CLI (Phase F) supports the same flow:

recipes publish --slug my-skill --dir ./skills/my-skill/

This packs the directory into a tarball, runs the local quality gate, and posts to POST /api/publish/request. The response is the same as the MCP tool.

Quality gate

The quality gate runs on every submission and checks for:

HIGH (blocks submission): curl | bash pipelines, hardcoded $HOME or username paths, binary executables in the tarball, path traversal in filenames, credential patterns (AWS keys, GitHub tokens, etc.), internal WiseChef infra references.
MEDIUM (warns, allows submission): External links not on the allowlist, use of sudo, large scripts (>500 lines), missing CHANGELOG section.
LOW (informational): Missing screenshots, short description (<100 chars), no template files.

Tier semantics

Set the tier field in your submission:

free — available to all API key holders, no subscription required. Good for open-source skills you want maximum adoption on. MIT or Apache-2.0 license preferred.
cook — available to Pro subscribers ($20/mo). This is the sweet spot for polished, maintained skills. Most catalog skills are cook-tier.
operator — available to Pro+ subscribers ($100/mo). For enterprise skills: fleet-specific tooling, high-compute workflows, bespoke integrations.

The reviewer may suggest a different tier if the skill's complexity or maintenance burden warrants it. Final tier is set at publish time.

Rate limits

recipes_publish_request is rate-limited to 1 submission per 24h per (user, slug) pair. This prevents flooding the review queue. If you need to resubmit quickly after a rejection, contact the reviewer on the GitHub issue.

Creator earnings

Once your skill is published, you earn a share of the subscription revenue attributed to its install and use events. Payouts are calculated monthly. See the Creator program for current revenue share percentages and payout minimums.

FAQ

Can I update a published skill?

Yes — increment the version in your SKILL.md frontmatter and call recipes_publish_request again with the updated content. Patch versions (e.g. 1.0.0 → 1.0.1) go through expedited review. Major versions (e.g. 1.x → 2.0.0) get a full review cycle.

What if the quality gate flags my skill?

HIGH severity findings block the submission — no issue is opened. Fix the finding and resubmit. MEDIUM and LOW findings are warnings: the issue is opened with the findings listed, and the reviewer may ask you to address them before approval.

How long does review take?

Standard review: we aim to respond within a few business days. Operator subscribers receive priority review. If your issue hasn't received a response, comment on it with a ping.

Publishing a skill.