Skip to content

Agent Discovery

Skillett is designed for AI agents to discover and use without manual configuration. Discovery works at two levels: static (skill files installed locally) and dynamic (runtime API queries).

When you run skillett login, the CLI installs a lightweight SKILL.md file in your agent’s skill directory:

---
name: skillett
description: Skillett agent skills platform — run `skillett help` to discover available commands.
user-invocable: false
---
# Skillett
Skillett is installed. Run CLI commands to discover and execute agent skills.
skillett help # see all commands
skillett skills # list available integrations
skillett run ... # execute an endpoint

This is the entry point. When an agent encounters this skill, it knows Skillett is available and can start issuing CLI commands.

The CLI auto-detects which agent platform you’re using by checking for platform-specific directories:

PlatformDetectionSkill path
Claude Code.claude/ exists.claude/skills/skillett/
Cursor.cursor/ exists.cursor/rules/skillett/
Windsurf.windsurf/ exists.windsurf/rules/skillett/
GenericNone of the above.skillett/skills/skillett/

When you download integration docs (skillett pull github), they’re written to a platform-appropriate path:

PlatformPull path
Claude Code.claude/skills/skillett-github/
Cursor.cursor/rules/skillett-github/
Windsurf.windsurf/rules/skillett-github/
Generic.skillett/skills/github/

Agents can discover available skills at runtime through the three-level API:

  1. skillett skills — What integrations are available?
  2. skillett skills github — What endpoints does GitHub have?
  3. skillett skills github create_issue — How do I use this specific endpoint?

This progressive disclosure means agents only fetch the documentation they need, reducing context window usage.

For agents that prefer local docs (or when network access is limited), skillett pull downloads the complete documentation for an integration:

Terminal window
skillett pull github

This writes the SKILL.md (overview, categories, workflows) and all endpoint docs to the local skill directory. The agent can then read these files directly without API calls.

A .well-known/agent.json endpoint is planned for automated agent discovery, following emerging standards for agent capability advertisement.