Local-first agent skill manager

Every agent skill on your machine, in one place.

Inventory every agent skill, find what's wrong, and manage it safely. SkillCat reads what you have installed across 90+ agents, shows the evidence behind every finding, and turns your own model into read-only optimization suggestions — fixes you can judge, never edits it makes on its own.

  • Free · MIT licensed
  • macOS + Windows
  • v1.2.0
  • Reads offline
  • 90+ agents
  • Read-only optimizer
  • No account needed
SKILL.md 0 problems
---
name: release-notes
description: Generate release notes from merged PRs
when_to_use: drafting a release, changelog
---

## When to Use
Use when drafting a release, or turning
merged pull requests into user-facing notes.
  • ~/.agents/skills
  • symlink ok
  • 1.2 KB · 1 file
  • sha 3f9a1c07

Works across 90+ known agent directories

  • Claude Code
  • Codex
  • Cursor
  • OpenCode
  • Gemini CLI
  • Windsurf
  • Trae
  • GitHub Copilot
  • Cline
  • Continue
  • Goose
  • Amp
  • Droid
  • Zed

01 · Skill management

Every skill, its real state, and one safe way to change it.

SkillCat reads the global, per-project and custom directories across 90+ agents, then shows the source, the lock metadata and the real link state of every skill — symlinked, dangling or drifted.

  • 01

    Inventory across 90+ agents

    Global, per-project and custom directories in one list, each with its source, size, update time and content hash.

  • 02

    The real link state

    Symlink, dangling link, copy drift or a local edit — the disk fact, not a guess, for every installed skill.

  • 03

    Findings with evidence

    Deterministic rules flag dangling links, missing lock records, copy drift, cross-scope shadowing and source conflicts, each with the evidence that triggered it.

  • 04

    Projects, auto-discovered

    Finds projects under your scan roots by their agent markers. Pin the ones you revisit; the registry stores paths only.

  • 05

    One safe install path

    Install, update or remove runs through the official skills CLI, streamed into a drawer with a confirmation step you can cancel.

SkillCat Skills view: installed skills on the left, and a detail panel on the right showing description, trigger profile, link state and body.
SkillCat Issues and scores view: findings on the left, colour-coded by severity, with the evidence and a suggested fix on the right.

Skills · real link state  ·  Issues & scores · evidence per finding

02 · Skill optimization

Ask your model for fixes. Never let it touch a file.

Turn your own model into a reviewer for one skill at a time. It returns a summary plus severity-ranked suggestions with before/after excerpts, grounded in a bundled rubric — and writes nothing back.

  • 01

    Severity-ranked, with evidence

    Every suggestion carries a title, a severity (high, medium or low), a rationale and optional before/after excerpts, so you judge the fix instead of trusting it.

  • 02

    Grounded in a bundled rubric

    Review criteria come from the built-in skill-optimizer skill. A same-named user skill is ignored, so a prompt injection in your skills cannot hijack the review.

  • 03

    Saved, not re-billed

    Results persist to optimizer.json, keyed to the skill and its content hash. Re-opening a review is free; only an explicit regenerate spends tokens again.

SkillCat skill detail page, Suggestions tab: a summary and severity-ranked optimization suggestions for one skill, each with current and suggested excerpts.

Suggestions tab · one skill at a time

Read-only guarantee No skill file is ever rewritten. The optimizer only returns suggestions — every change still goes through the official skills CLI, with an explicit confirmation step.

03 · Trigger statistics

See what actually fires — not just what could.

SkillCat extracts each skill's trigger profile, and can record which skills really fire inside your agent — with the trigger phrase, the task and the time. The stats stay on your machine.

  • 01

    Trigger profiling, positive and negative

    Extracts triggers from when_to_use, dispatch_intent, the description and “When to Use” sections. Annotate by hand in a sidecar; skill files are never rewritten.

  • 02

    Runtime activity, optional

    Install the OpenCode bridge to log real triggers: which skill, which phrase, which task, which session. Off by default, and a one-click uninstall.

  • 03

    Per-day and per-skill statistics

    Charts show triggers per day and the skills that fire most. A single sentence that fires several skills is aggregated as a conflict signal.

  • 04

    Private by default

    Trigger phrases are prompt fragments. They are stored only in your local config directory, never uploaded, with a retention window you control and a one-click clear.

SkillCat Usage log view: total triggers, skills fired, triggers per day and a table of recent triggers with the skill, phrase, task and agent.

Usage log · triggers and stats

What it does, and how far to trust it.

Every capability is labelled by how it knows what it knows. Disk facts are deterministic. Similarity results are heuristic and say so.

InventoryAcross 90+ known agents

Scans global, per-project and custom directories, then reports the source, the lock metadata and the real link state of every skill: symlinked, dangling, or a drifted copy.

core

ProjectsAuto-discovered

Finds projects under your scan roots by their agent markers. Pin the ones you revisit. The registry stores paths only, and removing an entry never deletes a file.

core

AnalysisDeterministic and heuristic

Deterministic rules read disk facts: dangling links, missing lock records, copy drift, local edits, cross-scope shadowing and source conflicts. Heuristics surface trigger overlap, negative contradictions and body duplication, each with a confidence score and the shared terms that caused it.

rules

Safe operationsOne install path

Every change streams progress into a drawer with a confirmation step and can be cancelled. Writes are delegated to the official skills CLI, so there is one install path, not two.

cli

Optimization suggestionsOptional, read-only

Ask your model to review a single skill and return read-only rewrite suggestions with before/after excerpts, grounded in a bundled reviewer skill. Nothing is written back; results are saved and kept until you regenerate them.

ai

Trigger profilingPositive and negative

Extracts what dispatches each skill from when_to_use, dispatch_intent, descriptions and “When to Use” sections. Annotate triggers by hand in a sidecar; skill files are never rewritten.

core

Runtime activityOptional, local-only

Install the OpenCode bridge to record which skills actually fire, with the trigger phrase, task and per-day and per-skill charts. Reversible in one click, and stored only in your local config directory.

optional

Remote searchskills.sh leaderboard

Browse the public index and the all-time, trending and hot leaderboards, read a SKILL.md before you commit, then install straight into the current scope.

online

AI evaluationOptional, your own key

Score every skill with your own model, then judge the pre-filtered duplicate and conflict candidates. Results are saved with a signature, so the UI can tell you when an evaluation has gone stale.

ai

Scan, analyze, act.

Three steps, and the third one is the only one that can change anything on disk.

01

Scan

Read the global and project lock files, enumerate the canonical directory, every known agent directory and your custom directories, then hash the content of each skill.

02

Analyze

Run the rule engine over every record, extract trigger profiles, compute similarity where it is cheap, and attach your saved annotations. Nothing is written back to a skill.

03

Act

Install, update or remove — each one confirmed, streamed and cancellable, and each one executed by the official skills CLI rather than by SkillCat itself.

Reads

Local and offline

Scanning uses the filesystem and lock files only. No network call is required to inventory or analyze your skills.

Writes

Delegated to the CLI

Every mutation runs through the official skills CLI. SkillCat never becomes a second source of truth, and never edits your files behind your back.

The rules it will not break.

P1

Reads are local and offline

Scanning uses the filesystem and lock files only. No network call is required to inventory or analyze your skills.

P2

Writes are delegated

Every install, update and removal runs through the official skills CLI. SkillCat never becomes a second source of truth.

P3

Analysis is explainable

Deterministic rules and heuristics are labelled separately, and every finding ships with the evidence that triggered it.

P4

AI is optional and explicit

Scoring, duplicate judgement and optimization suggestions run only when you ask, with your own key and endpoint. The optimizer is read-only — it never rewrites a skill. Nothing leaves your machine by default.

Download it, or build it.

Grab the latest build from the releases page, or clone the repo and run it in development mode.

macOS

Self-contained

  • Apple Silicon and Intel builds (dmg or zip).
  • The skills CLI is bundled — no Node.js install required.
  • Drag SkillCat into /Applications.
SkillCat-1.2.0-arm64.dmg

Windows

Needs Node.js

  • x64 NSIS installer.
  • Install / update uses your system Node.js, version 22 or newer.
  • The bundled CLI is macOS-only for now.
SkillCat-1.2.0-setup.exe
Unsigned builds The builds are not code-signed. On macOS, Gatekeeper blocks the first launch of a downloaded copy — right-click → Open, or clear the quarantine attribute: sudo xattr -dr com.apple.quarantine /Applications/SkillCat.app. On Windows, SmartScreen may warn about an unknown publisher; choose More info → Run anyway.

Questions worth asking.

Q1 How is this different from running the skills CLI on its own?

The CLI installs skills. SkillCat shows you the state of everything you have already installed: it reads what the CLI and your agents wrote to disk, explains each skill, and flags concrete problems before they bite.

Q2 Does it need an API key?

No. Inventory, trigger profiling and the rule engine all run without a model. An API key is only needed for the optional AI evaluation and the read-only optimization suggestions.

Q3 Which agents does it know about?

Over 90 known agent directories, including Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Windsurf and Trae. You can add your own skill directories as well.

Q4 Does it modify my skills?

Not silently. Scanning is read-only, and annotations live in a sidecar file in the config directory. Optimization suggestions are read-only too: they return before/after excerpts and never write to a skill. Any real change goes through the official skills CLI, with an explicit confirmation step.

Q5 Does it work offline?

Yes. Scanning, analysis and the usage log use only the filesystem and lock files, and the log stays local. Network access is limited to the optional remote search, leaderboard and AI evaluation.

Q6 Does it record what I type?

Only if you switch on runtime activity. With the optional OpenCode bridge installed, SkillCat logs which skill fired, the trigger phrase, the task and the time — stored locally in your config directory, never uploaded, with a retention window you set and a one-click clear. Uninstall the bridge and nothing is recorded.

Q7 Is it free?

Yes. SkillCat is MIT-licensed and open source, with no account, no subscription and no telemetry.

Q8 What do I need to run it?

macOS is self-contained: the skills CLI is bundled, so no Node.js install is required. On Windows, install Node.js 22 or newer.

Point it at your workspace and see what you have.

Free, open source, and read-only until you say otherwise — even the optimization suggestions.