Find reusable SKILL.md workflows by task, occupation or creator. SkillsMP links to public repositories and offers search APIs and metadata exports for deeper discovery.
Used this tool? Rate it
Used this tool? Rate it
SkillsMP is a discovery directory for agent skills sourced from public GitHub repositories. It organizes reusable task instructions and related files so people can explore what creators have packaged into SKILL.md, trace a listing to its repository, and find examples from other fields. The project describes itself as independent and states that it is not affiliated with Anthropic or OpenAI.
SkillsMP provides source discovery. Task capabilities belong to upstream skills and the assistants that use them; catalog inclusion does not establish that SkillsMP has performed a task or validated its result.
The service also offers REST keyword retrieval, shared search allowances across REST and MCP, and downloadable public skill metadata. These support finding and organizing references, rather than executing the indexed tasks.
A repository search locates a collection; an account search follows a creator; a skill search narrows the task. Occupation browsing provides an entry point when the starting question concerns a field of work rather than a known package.
Metadata exports contain identifiers, slug, title, author, repository, stars, update time, domain/category/occupation tags, summary and a source link. That combination preserves both a human-readable description and the route back to the original project. The exported data can support catalog organization, statistics or analysis performed by the recipient; SkillsMP does not provide that post-processing.
SkillsMP's discovery flow starts with the job the agent is meant to handle, then moves to actual source material through categories, creators or search.
SkillsMP highlights pstack as a collection of engineering workflows for investigation, implementation and verification, with poteto-mode as an entry point. This is a concrete example of using the directory to find how another project structures engineering work across stages, rather than searching only for an isolated code snippet.
The featured Marketing Skills collection by Corey Haines covers copywriting, SEO, conversion improvement and marketing campaigns. It illustrates a different use of the same directory: discovering how a creator packages recurring commercial tasks into reusable agent instructions.
SkillsMP's guide names Claude Code, Codex, ChatGPT and assistants that can use SKILL.md workflows as agent targets. These names describe the environments associated with the skill format, rather than a guarantee that one downloaded skill works unchanged in all of them.
Installation methods and compatibility vary by client and version. SkillsMP directs installation through the skill detail, source repository and the client's installation guide, and explicitly says not to assume every skill works on every platform. A platform name in the directory therefore needs to be read together with the particular package's files and requirements.
The REST search endpoint is GET /api/v1/skills/search, with the required query parameter q. Pagination starts at page 1; the default page size is 20 and the maximum is 50. Results can be sorted by stars or recent, with stars as the default. These are retrieval settings for catalog entries, not controls for running a skill.
Search also accepts a category slug, a SOC occupation slug and a detected content-language filter. Language codes include forms such as en, zh or ja; mul selects mixed-language content and und undetermined content. A content-language filter describes the indexed skill text, rather than translating the skill or proving that an assistant can execute it in that language.
Browsing and searching the SkillsMP directory are described as free in its public terms. Paid plans add metadata-export and authenticated search allowances; third-party services referenced by a skill are not covered by that statement of free directory access. Website browsing, API use and exports have separate access rules, minute limits and shared-capacity protection.
The following are the annual prices and allowances captured on October 22, 2026. The displayed monthly equivalents are not month-by-month payment amounts.
| Plan | Annual payment shown | Displayed monthly equivalent | Metadata-export allowance | Successful ordinary API/MCP searches |
|---|---|---|---|---|
| Free | US$0; no paid billing cycle | Not applicable | 1 successful task/day; up to 500 rows/task | 100/day |
| Standard | US$99 upfront/year | US$8.25/month | 20,000 new metadata entitlements/month; up to 10,000 rows/task | 5,000/month |
| Professional | US$299 upfront/year | US$24.92/month | 100,000 new metadata entitlements/month; up to 10,000 rows/task | 25,000/month |
Annual payment does not supply the entire year's allowance at once. Paid API and export allowances are still allocated monthly. Paid exports default to 1,000 rows per task, with the larger per-task maximum shown above; the monthly export meter counts newly granted metadata entitlements, so returned row count and chargeable new records can differ.
REST and MCP share the account's allowance for successful ordinary keyword searches, including when the subscription is annual. All account keys draw from that balance; rotating a key does not reset usage. Authenticated requests retain a separate limit of 30 attempts per minute.
Failed ordinary keyword searches do not consume the success allowance, while minute rate limits count attempts. Free successful export tasks and authenticated search allowances reset at 00:00 UTC. Paid API allowances operate by subscription month without a separate small daily budget.
Anonymous REST keyword search is documented at 50 requests per day and 10 per minute. Wildcard searches are unsupported. Anonymous MCP is described separately as a small hourly IP trial, so the anonymous REST figures should not be treated as its quota. Hybrid search also retains a separate pilot allowance rather than inheriting the ordinary keyword balance.
Paid exports deduplicate the same Skill ID on the same account for 30 days after its last successful delivery. Each successful repeat export restarts that window. Changing between CSV and JSONL or updating metadata does not consume more records within the window; once it expires, the record counts as new again. Free exports instead count successful tasks, and downloading an already generated file again does not create another task.
New export files are downloadable for 30 days after completion; existing files keep their recorded deadlines. File expiry does not return allowance used for successful delivery.
Renewal can be stopped through Account usage, with access continuing through the paid service period subject to the refund terms. The terms distinguish cancellation from requesting a refund or deleting an account. A checkout being created, or a browser returning from checkout, is not itself confirmation that payment was verified and the plan activated.
For the first purchase of an ordinary paid feature, the terms allow a full-refund request within 7 calendar days of the first successful charge, without a usage threshold. Advertising placements and sponsored displays are excluded. This first-purchase guarantee does not restart at renewal; outside it, cancellation alone normally does not refund a period that has already begun. Billing errors, non-delivery and mandatory refund rights remain exceptions.
Skills.sh combines a public directory with the open-source skills CLI. Its documentation gives npx skills add vercel-labs/agent-skills as an installation example. That makes it a relevant comparison when the desired entry point is a skill-installation command, while SkillsMP's documented strengths here are source exploration, occupation-based discovery, search access and metadata exports.
The ranking signal also differs from a source catalog. Skills.sh says its leaderboard uses anonymous aggregated installation telemetry from its CLI. That is an install-based popularity signal, whereas SkillsMP's search interface offers stars and recent updates and its cross-field browsing emphasizes source context. The metrics answer different questions and are not interchangeable measures of reliability.
Skills.sh also states that it cannot guarantee the quality or security of every listed skill. An install-based ranking is not a safety certificate.
An independent study of 17,022 skills sampled from a 170,226-skill SkillsMP corpus identified 520 affected skills with 1,708 credential-security issues. The study used static analysis, sandbox testing with mock credentials and manual examination. Its result concerns that sampled corpus, rather than every current listing, a present-day failure rate or SkillsMP's own handling of website account data.
Each indexed skill remains subject to its original repository's license, and SkillsMP does not claim ownership of indexed skills. Permission to browse or download through the directory is separate from the repository's conditions for using its content. A directory description does not replace the license governing that particular source.
The terms permit personal and commercial browsing, searching and downloading while requiring compliance with each repository's license.
SkillsMP's privacy policy discloses analytics collection of page views, clicks, search queries and browsing behavior, along with browser, device, operating-system and IP information.
Theme preferences and browsing history are stored locally in the browser. Google or GitHub sign-in introduces account identifiers, email and available profile information into cloud authentication, while signed-in favorites are linked to the account in a cloud database. Local browsing storage therefore does not mean every account-related record stays on the device.
API and export activity also has an account record: API credentials and key timestamps, usage periods and totals, export filters, file formats, task status, accounting events and delivered Skill identifiers. Billing records include purchase and renewal state and refund information. These serve account operation, allowance accounting, reconciliation and support; export files are only one part of that record set.
Expiry blocks export downloads, and cleanup is retried; the policy does not guarantee physical deletion at the exact deadline. Export expiry also does not erase task metadata, usage summaries, subscription orders or payment and refund records. The download period should therefore be read as an access deadline, rather than an account-wide retention period.
Clearing browser data, deleting an API key, stopping renewal or allowing an export file to expire does not erase all cloud records. The policy offers contact-based access, correction and deletion requests alongside browser-data and account controls.
