Agent Integration
SiteSkills skills are agent-agnostic. A skill is a set of structured Markdown instructions -- any agent that can read files and control a browser can use them.
Supported Agents
| Agent | Type | Install Path | Notes |
|---|---|---|---|
| Claude Cowork | Native SKILL.md | .claude/skills/ | Auto-discovers skills from directory |
| Browser Use | Python framework | ./skills/ | Pass skills directory to agent config |
| Stagehand | TypeScript framework | Configurable | Read SKILL.md for navigation guidance |
| Playwright MCP | MCP server | Configurable | Any MCP-compatible agent can use skills |
How It Works
- Install a skill using the CLI:
siteskills install <skill-name> - The CLI places the skill directory in the correct location for your agent
- The agent reads
SKILL.mdwhen a matching trigger phrase is detected - The agent follows the navigation steps, using DOM hints and fallbacks
- Execution results can be reported back via
siteskills report
Skill Discovery
Agents discover skills through different mechanisms:
- Claude Cowork: Scans
.claude/skills/on startup, matches triggers from SKILL.md frontmatter - Browser Use: Loads from configured
skills_dir, matches by domain and trigger - Stagehand: Skills are loaded programmatically or referenced in agent configuration
Custom Agent Integration
If you are building your own agent, here is how to consume SiteSkills skills:
- Download skills via the API or CLI
- Parse the SKILL.md frontmatter for
name,description, andtriggers - When a user request matches a trigger, load the full SKILL.md
- Follow the
Navigation Flowsteps sequentially - Use
DOM hintsto find elements (try selectors in priority order) - Respect
Wait conditiontimeouts before proceeding - On failure, follow the
Fallbackinstructions for each step - After completion, check the
Verificationsection to confirm success
Execution Reporting
Report execution results to improve community success rates:
bash
siteskills report <skill-name> --success --time 12400 --agent my-agentOr via the API:
POST /api/telemetry/executionSee API Reference for the full telemetry payload.