webmcp
GitHub提供WebMCP在Web项目中的集成指南，支持通过JS或HTML声明式API为AI代理添加结构化工具。涵盖初始化配置、工具开发、调试、审计及测试全流程。
触发场景
安装
npx skills add qdhenry/Claude-Command-Suite --skill webmcp -g -y
SKILL.md
Frontmatter
{
"name": "webmcp",
"description": "Implement WebMCP in web projects — add browser-native structured tools for AI agents using imperative JS or declarative HTML APIs. Full lifecycle from setup through testing and optimization."
}
<essential_principles>
WebMCP is a browser-native standard that exposes structured tools for AI agents on websites. Instead of screen-scraping, agents interact through typed JavaScript APIs and HTML annotations.
Two APIs:
- Imperative API —
window.navigator.modelContext— register tools via JavaScript - Declarative API — HTML
toolname/tooldescriptionattributes on<form>elements
Prerequisites: Chrome 146.0.7672.0+, chrome://flags/#enable-webmcp-testing enabled.
Core principles that always apply:
-
Tools must be atomic and composable — one function per tool, no overlapping tools with nuanced differences. Combine similar tools into one with input parameters.
-
Accept raw user input — tools should accept natural strings (e.g., "11:00") not computed values (e.g., minutes-from-midnight). Minimize cognitive computing for the model.
-
Validate in code, not schema — schema constraints are helpful but not guaranteed. Validate within execute functions and return descriptive errors so agents can self-correct.
-
Update UI before returning — agents use UI state to verify execution. Ensure execute/submit logic updates visible state before resolving.
-
Positive, verb-first descriptions — describe what the tool does and when to use it. Avoid negations. "Creates a calendar event for a specific date" not "Do not use for weather."
</essential_principles>
- Set up WebMCP in a project (initial integration)
- Add a tool using the imperative JavaScript API
- Add a tool using the declarative HTML form API
- Debug WebMCP tools that aren't working
- Audit existing WebMCP implementation for best practices
- Test WebMCP tools
- Something else
Wait for response before proceeding.
After reading the workflow, follow it exactly.
<reference_index>
All in references/:
APIs: imperative-api.md, declarative-api.md Design: tool-design.md Events: events-and-css.md Quality: testing.md, anti-patterns.md
</reference_index>
<workflows_index>
All in workflows/:
| Workflow | Purpose |
|---|---|
| setup-webmcp.md | Initial WebMCP integration into a project |
| add-imperative-tool.md | Register tools via JavaScript API |
| add-declarative-tool.md | Annotate HTML forms as WebMCP tools |
| debug-webmcp.md | Diagnose and fix WebMCP issues |
| audit-webmcp.md | Review implementation against best practices |
| test-webmcp.md | Test tools with inspector extension and agents |
</workflows_index>
<templates_index>
All in templates/:
| Template | Purpose |
|---|---|
| imperative-tool.md | Scaffolding for JS-registered tools |
| declarative-form.md | Scaffolding for HTML form-based tools |
</templates_index>
<verification_loop>
After every WebMCP change:
- Open Chrome DevTools console — check for JS errors
- Verify
navigator.modelContextexists (Chrome flag enabled?) - Use Model Context Tool Inspector extension to list registered tools
- Manually execute a tool from the inspector to confirm it works
- Check that UI updates are visible after tool execution
Report to user:
- "Tools registered: X imperative, Y declarative"
- "Tool [name]: executes successfully / returns error [detail]"
- "UI state: updates correctly after execution / missing update"
</verification_loop>
版本历史
- e89b2f0 当前 2026-08-20 06:58
