Sign inSet up
Docs

Connect, write, compose.

Quickstart

Three steps, about a minute. You need a Skillfiles account and an agent that supports MCP with OAuth sign-in. There are no API keys: the agent signs in for itself.

1

Add the library

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

Approve in your browser

$ claude mcp login skillfiles

Sign in, check the agent named is the one you added, and click Approve.

3

Ask

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

In Claude Code you can also type /mcp inside a session, choose skillfiles and pick Authenticate. For any other agent, add an MCP server of type Streamable HTTP at https://app.skillfiles.dev/mcp; a client that supports MCP OAuth signs in the same way.

Once approved, the agent holds an access token that lasts 15 minutes and renews it by itself, so you don’t sign in again. To see your connected agents or cut one off, open the connections page.

Concepts

Four ideas cover everything else in these docs.

Library

All your skills. Each account has one library, reached at one MCP address, and an agent you approve can reach only yours.

Skill

A folder in the open Agent Skills format, with a SKILL.md and any supporting files. Search matches on its name and description.

Connected agent

An MCP client you approved in your browser. It acts for you, and you can revoke it at any time.

Review

What agents write goes live at once and waits on the dashboard until you approve or reject that exact version.

MCP tools

The tools Skillfiles gives your agents. You don’t call them yourself: you ask in ordinary language, and the agent picks the tool.

search_skills

Finds the skills that fit a task. Keyword and meaning search find candidates from the query; then Claude Haiku reads the query, the intent and each candidate, leaves out the ones that don’t fit and orders the rest. Each result has a confidence of strong possible or weak, and no_good_match is true when nothing fits. If the judging step can’t run, results come back unjudged with judged: false.

queryOne line describing the kind of work, worded like a skill description.intentOptional. The agent’s fuller account of the work, which the judge reads against each candidate.repoOptional. The output of git remote get-url origin. Skills scoped to this repository come first; skills scoped to others are left out.limitOptional. How many results, from 1 to 20. The default is 5.

read_skill

Reads a skill’s SKILL.md and lists its other files, with the skill’s version and a link to the whole folder as an archive, valid for five minutes. With that version, it reads a supporting file.

uriThe skill’s address, for example skill://pr-review/SKILL.md.versionRequired for a supporting file, so it always matches the SKILL.md that was read.

write_skill

Creates a skill, or replaces one completely. The write takes effect at once and waits for your review.

nameLower-case letters, digits and single hyphens, matching the name in SKILL.md.filesEvery file of the skill, each with content or content_base64. A file left out is removed.reasonRequired. One or two sentences on why the agent is writing, for you to read when you review it.expected_versionLeft out when creating. When changing a skill, the version read_skill returned; if the skill changed in the meantime the write is refused, so nothing is overwritten by accident.

delete_skill

Deletes a skill and all its files.

nameThe skill to delete.expected_versionIts current version, so a skill that changed after it was read is never deleted.

revoke_skill

Takes back the agent’s own write before you approve it: a new skill is deleted, and a changed one goes back to the version you approved.

nameThe skill the agent wrote.versionThe version write_skill returned.

Agent commands

Skillfiles has no command-line tool of its own. You add it, and sign in, with your agent’s own MCP commands.

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

Adds the library to Claude Code for every project on this machine.

$ claude mcp login skillfiles

Opens your browser to approve Claude Code.

$ codex mcp add skillfiles --url https://app.skillfiles.dev/mcp

Adds the library to Codex.

$ codex mcp login skillfiles

Opens your browser to approve Codex.

Skill file format

A skill is a folder of files in the open Agent Skills format. SKILL.md is Markdown with a short header. Search matches on the name and description, so make the description say what the skill is for and, for a narrow skill, what it isn’t for.

pr-review/SKILL.md
---
name: pr-review
description: How I review any pull request before merging, in any repository.
---
# PR review
Load the commit-messages and test-style skills first.
Then check each item in references/checklist.md.
  • name is lower-case letters, digits and single hyphens, and is the skill’s name in the library. skillfiles is reserved.
  • description is required. It is what search and the judge read first.
  • A skill can hold up to 512 files and 16 MiB, and one write_skill call can carry up to 4 MiB.
  • Supporting files can be text or binary. The agent reads them with read_skill when the skill refers to them.

Bootstrap skill

The bootstrap skill tells an agent to search your library before it starts a task, how to choose from the results and how to follow a library skill. Agents already get the same rule from the MCP server’s instructions when they connect; installing the skill makes it stronger. Ask your agent once:

you → agent
Read skill://skillfiles/SKILL.md from the skillfiles MCP server and install it as a personal skill called skillfiles.

Skills that name skills

Keep skills small and let one skill name the others a task needs, in plain words: “Load the commit-messages skill first.” The agent reads each named skill with read_skill and follows it, and does the same for any skill that one names. A path such as skills/commit-messages/SKILL.md written on another machine is read as the library address skill://commit-messages/SKILL.md, never as a local file.

Repository scope

A skill that belongs to particular repositories names their git remotes in metadata.repo, separated by commas. Searches made from one of those repositories rank it first; searches from anywhere else leave it out. Every way of writing a remote compares equal, so git@github.com:acme/billing.git and https://github.com/acme/billing match.

billing-db-migrations/SKILL.md
---
name: billing-db-migrations
description: Writes database migrations for the billing service. Not for other services.
metadata:
  repo: github.com/acme/billing, github.com/acme/ledger
---

Reviewing agent writes

A skill an agent adds or changes goes live at once, marked as needing your review. Open the dashboard and choose Review: each change shows the reason the agent gave, a diff against the version you approved and flags from fixed rules, such as an added script, a widened repository scope or a line that pipes a download into a shell. Rejecting a new skill removes it at once; rejecting a change puts back the version you approved.

Everything works from the keyboard: j and k move between changes, a approves, r rejects, z undoes and ? lists the rest.

GitHub backup

The library can keep a copy of every skill in a GitHub repository. Every hour it pushes one commit with whatever changed, and the connections page has Back up now and shows the last run. The repository is laid out like any repository with local skills, with each skill under .agents/skills/<name>/, so agents can use it without Skillfiles. It is a mirror: change skills through the library, because hand edits in the repository are put back on the next run.

Backup is part of Pro. [How to connect your own repository, once the GitHub App is available]