Skills on inference shell follow the Agent Skills open standard. This means skills you create here work with Claude Code, Cursor, Windsurf, and any tool that supports the spec.
The Skill URL
Every skill has a stable content URL:
1https://api.inference.sh/skills/{namespace}/{name}/contentThis URL serves the skill's SKILL.md content (frontmatter + instructions). It can be used directly in any agent config that accepts a URL for context or instructions.
Supporting files are available at:
1https://api.inference.sh/skills/{namespace}/{name}/files/{path}For example: .../files/references/api-docs.md, .../files/scripts/helper.py
Claude Code
Claude Code discovers skills from the local filesystem. To use an inference shell skill:
Manual Install
Download the skill files and place them in Claude Code's skills directory:
1~/.claude/skills/{skill-name}/2├── SKILL.md3├── references/4│ └── api-docs.md5└── scripts/6 └── helper.pyThe download endpoint provides a manifest of all files with their URLs:
1GET https://api.inference.sh/skills/{namespace}/{name}/downloadReturns:
1{2 "name": "code-review",3 "namespace": "inferencesh",4 "version": "1.0.0",5 "files": [6 { "path": "SKILL.md", "uri": "https://cdn.inference.sh/..." },7 { "path": "references/checklist.md", "uri": "https://cdn.inference.sh/..." }8 ]9}Download each file from its uri and write to the corresponding path under ~/.claude/skills/{name}/.
Belt CLI
1belt skill use inferencesh/code-reviewPrints the skill's main instructions (SKILL.md body) to stdout. Public skills do not require authentication.
When a skill has supplementary files (references/, scripts/, examples/), the CLI prints tips after the instructions:
1 tip: this skill has 3 supplementary file(s) not included above.2 install locally: belt skill add inferencesh/code-review3 list files: belt skill files inferencesh/code-reviewbelt skill add— downloads all files into your local skills directory (for Claude Code, Cursor, etc.)belt skill files— lists supporting file paths without downloading
Belt plugin install (all supported agents)
One command wires belt into every coding agent on the machine:
1belt login2belt plugin install # detects installed agents, sets up all of them3belt plugin install claude # or name them: claude codex copilot cursor droid gemini goose4 # grok hermes kilo kimi kiro omp opencode pi qwen windsurf5belt plugin doctor # verify hooks, skills, and rulesEach install writes three things:
- Lifecycle hooks in the agent's own hook format:
belt suggest --jsonon prompt submit,belt reviewon stop and before compaction. Claude Code and Codex install these through their plugin marketplaces (belt-sh/skills); the rest get a hooks config written directly. - Five core skills (
suggest,skill,apps,remember,belt) into the agent's skills directory. Skills are pulled from the platform when you are logged in; embedded copies are used offline. - A belt section in the agent's instruction file, the file it loads into its system prompt:
~/.claude/CLAUDE.md,~/.codex/AGENTS.md,~/.config/opencode/AGENTS.md,~/.gemini/GEMINI.md,~/.kiro/steering/belt.md,~/.copilot/instructions/belt.instructions.md, and so on. It sits betweenandmarkers so it can be updated or removed without touching your own notes. Cursor has no global rules file on disk, so at user scope it gets hooks and skills only.
--scope project writes into the current repo instead (AGENTS.md, CLAUDE.md, GEMINI.md, .cursor/rules/belt.mdc, .github/instructions/belt.instructions.md, .kiro/steering/belt.md). --rules=false skips the instruction file. belt plugin uninstall reverses all of it.
After installing for Codex, run /hooks in Codex to review and trust belt hooks. If you previously had belt hooks in ~/.codex/hooks.json from a manual or fallback install, the plugin install removes those entries automatically so hooks do not run twice. Restart the other agents after install.
All supported agents
belt plugin install covers 17 agents with the same three-part install. Claude Code and Codex go through their plugin marketplaces (Codex falls back to a direct install when the codex CLI is missing). Everything else gets the five core belt skills (suggest, skill, apps, remember, belt), a hooks config in the agent's own format, and the belt section in the agent's instruction file. Skills are downloaded from the platform when you are logged in; embedded copies are used offline.
Hooks run belt suggest --json on prompt submit and belt review --agent <name> on stop and before compaction. Claude Code and Codex plugin installs also run belt plugin hook session-start, which starts device authorization when no API key is configured.
| Agent | Skills directory | Hooks config | Instruction file |
|---|---|---|---|
| claude | ~/.claude/skills/ (plugin) | ~/.claude/settings.json (plugin) | ~/.claude/CLAUDE.md |
| codex | ~/.agents/skills/ | ~/.codex/hooks.json (plugin when codex is on PATH) | ~/.codex/AGENTS.md |
| copilot | ~/.copilot/skills/ | ~/.copilot/hooks/belt.json | ~/.copilot/instructions/belt.instructions.md |
| cursor | ~/.cursor/skills/ | ~/.cursor/hooks.json | — (Cursor Settings → Rules) |
| droid | ~/.factory/skills/ | ~/.factory/hooks.json | ~/.factory/AGENTS.md |
| gemini | ~/.gemini/skills/ | ~/.gemini/settings.json | ~/.gemini/GEMINI.md |
| goose | ~/.config/goose/skills/ | ~/.agents/plugins/belt-test/hooks/hooks.json | ~/.config/goose/.goosehints |
| grok | ~/.grok/skills/ | ~/.grok/hooks/belt.json | ~/.grok/AGENTS.md |
| hermes | ~/.hermes/skills/ | ~/.hermes/config.yaml | — (SOUL.md is identity only) |
| kilo | ~/.kilocode/skills/ | ~/.kilo/plugins/belt.ts | ~/.kilocode/rules/belt.md |
| kimi | ~/.kimi-code/skills/ | ~/.kimi-code/config.toml | ~/.kimi-code/AGENTS.md |
| kiro | ~/.kiro/skills/ | ~/.kiro/agents/belt.json, selected with chat.defaultAgent | ~/.kiro/steering/belt.md |
| omp | ~/.omp/agent/skills/ | ~/.omp/agent/extensions/belt.ts | ~/.omp/agent/AGENTS.md |
| opencode | ~/.config/opencode/skills/ | ~/.config/opencode/plugins/belt.ts | ~/.config/opencode/AGENTS.md |
| pi | ~/.pi/agent/skills/ | ~/.pi/agent/extensions/belt.ts | ~/.pi/agent/AGENTS.md |
| qwen | ~/.qwen/skills/ | ~/.qwen/settings.json | ~/.qwen/QWEN.md |
| windsurf | ~/.codeium/windsurf/skills/ | ~/.codeium/windsurf/hooks.json | ~/.codeium/windsurf/memories/global_rules.md |
See Coding agents for full setup details, including project-scope paths.
Manual install
Any agent that supports the Agent Skills standard can use inference shell skills without belt plugin install:
- Get the skill files via the download endpoint or skill URL
- Place them where the agent expects to find skills
- The agent discovers and uses them like any other skill
Check your agent's documentation for where skills are stored and how they're discovered.
URL-Based Usage
Some agents support loading skills directly from a URL rather than the filesystem. In that case, point to the content URL:
1https://api.inference.sh/skills/{namespace}/{name}/contentThis returns the full SKILL.md content — frontmatter + instructions — ready for the agent to parse.
The Roundtrip Guarantee
Skills created on inference shell produce the exact same SKILL.md format as the open standard. A skill created in the web editor and installed locally is indistinguishable from one written by hand:
- All spec frontmatter fields:
name,description,license,compatibility,metadata,allowed-tools - Same markdown body
- Same directory structure (
references/,scripts/,examples/,templates/)
This is by design — inference shell is a creation and hosting platform, not a proprietary format.
Next
- Creating skills — build skills on inference shell
- The registry — browse and share public skills
- Agent Skills standard — the open spec