The right skill, and only the right skill.
Skillfiles is a private skill library your agents reach over MCP. They search it, read the skills that fit the task, and leave the rest alone.
Agents load a skill, not a library.
Each task starts with a search. The agent describes the work, Skillfiles finds candidates and Claude Haiku judges each one against that description. The agent reads only the skills that clearly fit.
- search_skills returns candidates marked strong, possible or weak
- read_skill returns one skill’s
SKILL.mdand a list of its other files - The agent’s context stays small however large the library grows
One skill read. The rest of the library never enters the agent’s context.
Small skills, composed.
A skill can tell the agent to load other skills by name. The agent reads each one it is pointed to with read_skill, and follows any they name in turn.
- Write a method once and name it from every skill that needs it
- Named skills don’t count towards the two skills an agent loads from one search
- Skills that nothing names stay unread
Skills that belong to one codebase.
Set metadata.repo in a skill’s frontmatter to the git remotes it belongs to. Searches from those repositories rank it first, and searches from anywhere else leave it out.
- One remote, or several separated by commas
- SSH and HTTPS forms of the same remote match
- Skills without
metadata.repoapply everywhere
Returned only when the agent works in:
Every change kept, and checked by you.
Every write makes a new version, named by its contents. What an agent writes goes live at once and waits on the dashboard’s Review screen until you approve it. Rejecting a change puts back the version you approved.
- The agent’s reason, a diff and flags for risky lines on every change
- Approve, reject and undo from the keyboard
- On Pro, an hourly backup to a GitHub repository you own, laid out so agents can use it without Skillfiles
9f3c1a…51be02…0a77d4…See what your agents actually do.
The dashboard records every search, read, write and delete, so you can see which skills earn their place and which searches come back empty.
- Tool calls over a period you choose, against the period before
- Every skill with how often it was read, including the ones never read
- Every search with its results and what the agent read next
- Which agents were active, and what each one did
If the library is down, work continues.
When an agent can’t reach Skillfiles, it carries on without your skills and tells you so in one line. If it only needs you to sign in again, it stops and asks instead.
- No outage of ours blocks your work
- The agent says when it worked without the library
- Your GitHub backup works without Skillfiles
Bring the agents you already use.
Anything that speaks MCP and signs in with OAuth can connect. You approve each one once in your browser.
Write a skill once.
Every agent can find it.
Connect Claude Code, Codex or any OAuth MCP client to one library.