Give your coding agent a mission control for every skill you install.

Languages: English | 简体中文 | 日本語 | 한국어 | Español | Français

Your coding agent can design UI, test in a browser, deploy apps, review security, read docs, build CLIs, run notebooks, and call MCP tools. The problem is simple: the more powers you install, the easier it is for the agent to miss the right one.

SkillOS fixes that. It helps the agent notice what is installed, understand what each capability is for, choose the right chain of skills, explain the choice, and keep risky actions under your safety rules.

You do not need to remember whether the right tool is called playwright, security-threat-model, vercel-deploy, jupyter-notebook, or something else. Say what you want. SkillOS helps the agent translate that request into a real workflow.

SkillOS turns a folder full of skills into an agent that feels prepared.

• For beginners: ask in plain language. The agent should figure out the skill names.
• For builders: connect skills, MCP tools, rules, plugins, hooks, and client presets without hardcoding every case.
• For teams: keep decisions local, explainable, testable, and safer by default.
• For new skills: install them once, then let inventory and routing discover them next time.

You say:

I do not know UI design. Make this dashboard look professional and check it.

SkillOS helps the agent think:

This is not one task.
It needs UI judgment, implementation, browser verification,
screenshot QA, and maybe a security check before delivery.

Then it recommends a staged chain instead of guessing one random tool.

Another example:

Deploy this app, but do not break anything.

SkillOS can inspect the repo, look for installed deployment skills, add verification steps, flag credential risk, and explain what it selected or skipped.

The agent market is moving fast. Codex, Claude Code, Cursor, Windsurf, OpenHands, OpenClaw, MCP, and the Agent Skills ecosystem are all adding ways to package specialized capabilities. That is powerful, but it creates a new problem: someone has to decide which capability to use, when to use it, and how to combine it with the next step.

SkillOS is built for that missing layer.

Read the market notes in docs/market-context.md, then read the full product explanation in docs/product-overview.md.

SkillOS has two install layers:

• Agent Skill / plugin entry: teaches your agent when and how to call SkillOS.
• Local runtime: installs the skillos CLI and skillos-mcp-server.

For the best experience, install both.

More details: docs/distribution.md and docs/installation.md.

SkillOS is prepared for multiple developer discovery channels, not just GitHub:

• Agent Skills / skills.sh / GitHub Copilot skills.
• Claude Code plugin marketplace.
• GitHub Release zip downloads.
• npm preview packages for @skillos/core, @skillos/adapters, @skillos/mcp-server, and @skillos/cli.
• MCP Registry, Glama, Smithery, and PulseMCP.
• OpenClaw / ClawHub.
• Awesome lists and community launch channels.

Publishing details: docs/publishing-platforms.md, docs/launch-checklist.md, docs/mcp-distribution.md, and docs/community-launch-kit.md.

Install the agent-facing skill:

npx skills add xiaoxiaofeiya/SkillOS -g

Install the local runtime on Windows:

irm https://raw.githubusercontent.com/xiaoxiaofeiya/SkillOS/main/scripts/install-from-github.ps1 | iex

Install the local runtime on macOS/Linux:

curl -fsSL https://raw.githubusercontent.com/xiaoxiaofeiya/SkillOS/main/scripts/install-from-github.sh | bash

Then run:

skillos doctor
skillos setup --safety approve
skillos inventory
skillos recommend "Make this UI professional and verify it in a browser"
skillos explain --last

On Windows PowerShell, use skillos.cmd instead of skillos if execution policy blocks npm-generated .ps1 shims:

skillos.cmd doctor

Use --format json on any command when another agent or script needs stable machine-readable output.

This repository includes an isolated install verifier:

npm run verify:install

It tests the local Agent Skills path, local runtime install, linked commands with a temporary npm prefix, zip extraction, CLI smoke tests, and Claude marketplace metadata without writing to your real global skill or client configuration.

Network-dependent checks are separate:

npm run verify:install:network

If GitHub clone or remote npx skills add xiaoxiaofeiya/SkillOS fails because of a reset, timeout, proxy, or TLS issue, use the source install or the release zip path below. The runtime and agent-facing skill are separate layers; npx skills add installs the agent entry, while the runtime installer provides the skillos and skillos-mcp-server commands.

git clone https://github.com/xiaoxiaofeiya/SkillOS.git
cd SkillOS
powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1

git clone https://github.com/xiaoxiaofeiya/SkillOS.git
cd SkillOS
bash scripts/install.sh

The installer runs npm install, builds the workspace, links the skillos and skillos-mcp-server commands, and generates local client presets with skillos setup --safety approve.

Run without global command linking:

node scripts/install-local.mjs --no-link

skillos setup initializes local .skillos/ data when needed, renders client presets, writes generated presets to .skillos/generated-presets/, and shows planned client config changes.

It does not apply real client config files by default.

Review a client preset:

skillos preset diff --client codex

Apply only after review:

skillos preset apply --client codex --confirm

skillos doctor
skillos inventory
skillos recommend "Deploy this app safely"
skillos explain --last
skillos feedback --decision <id> --prefer playwright --avoid vercel-deploy
skillos presets --out dist-presets
skillos pack verify

• @skillos/core: inventory, capability cards, routing, repo signals, safety config, decision logs, feedback memory, and evals.
• @skillos/adapters: Codex, Claude Code, Cursor, Windsurf, OpenHands, OpenClaw, and OpenClaw-like presets.
• @skillos/mcp-server: JSON-RPC stdio MCP server exposing SkillOS tools.
• @skillos/cli: CLI for setup, init, inventory, recommend, inspect, eval, doctor, preset diff/apply, feedback, and package verification.

The MCP server exposes:

• inventory_skills
• search_skills
• inspect_skill
• recommend_skill_chain
• render_skill_context
• record_decision
• record_feedback
• run_eval
• detect_skill_gap
• render_client_preset
• verify_client_setup
• explain_decision

Tool results use a stable envelope:

{ "ok": true, "version": 1, "data": {} }

Errors use:

{ "ok": false, "version": 1, "error": { "code": "example", "message": "..." } }

SkillOS stores data locally by default and does not upload telemetry.

Safety profiles:

• suggest: recommend only.
• approve: require confirmation for medium/high-risk work.
• auto: allow low/medium-risk work and block high-risk work unless a client policy overrides it.

High-risk actions include deployment, credential use, external private-data sends, auth/security changes, and destructive operations.

Local routing is the default. If modelEnhancement.enabled is true and the configured API key environment variable exists, SkillOS can call OpenAI-compatible endpoints for embedding rerank, optional LLM rerank, skill summaries, and failure-review suggestions.

If credentials are missing or a request fails, SkillOS falls back to local routing and records the reason.

Build and verify:

npm run release:local

The zip package excludes .skillos/, local logs, credentials, node_modules, TypeScript build-info files, and local path metadata. release:local also runs isolated installation verification.

GitHub Actions builds and uploads dist/skillos.zip as a workflow artifact. Tags matching v* create a GitHub release with the zip attached.

Release details: docs/release.md.