Sign inSet up
Skills over MCP

One skill library for every agent.

Claude Code, Codex and other MCP clients search one library and read only the skills a task needs. Write a skill once. Every agent you connect can find it.

MCPhttps://app.skillfiles.dev/mcp
Works withClaude CodeCodexAny OAuth MCP client
your agent→skillfiles://live
1
Requestyou → agent
Write a commit message for these changes
2
Searchsearch_skills
commit-messagesstrong
pr-reviewpossible
release-notesdropped
2 kept · 1 dropped
3
Chooseread_skill
strongest matchskill://commit-messages/SKILL.md
4
Readskillfiles → agent
commit-messages/SKILL.mdreturned
---name: commit-messagesdescription: Write commit messages for staged changes---# Commit messagesSummarise the change in one line, then explain why.
How your agent finds a skill

Search first. Read only what fits.

Your agent never loads the whole library. It asks for skills that match the request, keeps the strong ones and reads a single file.

01

Search

The agent calls search_skills with what you asked for.

02

Propose

Skillfiles rates each candidate strong, possible or weak, and drops what doesn’t fit.

03

Choose

The agent picks the strongest match and calls read_skill.

04

Read

Skillfiles returns that skill’s SKILL.md, and nothing else.

Skills that build on skills

Small skills, composed into one request.

A skill can name other skills, and the agent reads those too. Keep each skill small, then let one skill pull in the chain it needs.

library·13 skills
Request
Ship the billing fix
13
skills in the library
10
read for one request

For this request the agent reads ship-feature, which names pr-review, release-notes and deploy-staging, and they name the rest. 10 skills read for one request. incident-notes, on-call and hotfix stay unread.

One owner, many agents

Your library. Every agent acts for you.

One account owns the library. Each agent you connect signs in once and works on that account’s behalf.

Claude CodeCodexMCP clientSkillfilesyour libraryyouowner

Each line is an OAuth sign-in you approved once. One address, and no API keys.

Whose library is it?
Yours alone. One account owns the library, and every agent you connect acts for you.
Which agents can connect?
Claude Code, Codex and any MCP client that signs in with OAuth. There are no keys to copy.
How do I move the skills I already have?
Ask your agent to save each folder from its skills directory to the library. They arrive on the dashboard’s Review screen like anything else it writes.
Can a skill belong to one repository?
Yes. Name its git remotes in metadata.repo, and searches from other repositories leave it out.
Where are my skills kept?
In your library, versioned by their contents. It can also mirror them to a GitHub repository every hour, laid out as an ordinary skills folder, so your agents can still use them without Skillfiles.
What if the library is down?
Your agent carries on without it and tells you so.
One address, no keys

Connect an agent in a minute.

Agents sign in through OAuth. You approve each one once in your browser, and there are no API keys to paste or rotate.

1

Add the library

$ claude mcp add --transport http --scope user skillfiles https://app.skillfiles.dev/mcp
2

Approve in your browser

Sign in once, check the agent named is the one you added, and approve it.

$ claude mcp login skillfiles
3

Ask in ordinary language

you → agent
Save this checklist as a skill called pr-review.
write_skillpr-review/SKILL.md

Pricing

Free for a small library. Pro when it grows.

Free covers 25 skills and 50 MB of files. Pro is $10 a month for 2,000 skills and 2 GB of files, more than a year of usage history and an hourly GitHub backup.

Compare plans

Write a skill once.
Every agent can find it.

Connect Claude Code, Codex or any OAuth MCP client to one library.