Skip to content

Repository files navigation

webmcp

Get your website ready for AI agents: audit, implement, verify.

WebMCP lets a web page register typed tools that in-browser AI agents call directly. Agents act through your site's own logic instead of scraping the DOM. This plugin teaches your coding agent (Claude Code, Codex, or any skills.sh host) to audit your site, propose the right tools, implement them on the standard document.modelContext API, and verify them in a real browser.

Built by Ora, the agent-readiness ranking. No login, no hosted service, no telemetry. Your code stays on your machine. Generated page tools target only the standard API; the optional bridge strategy uses a small MIT library, @ora-ai/webmcp-bridge.

Install

Claude Code

/plugin marketplace add ora/webmcp
/plugin install webmcp@ora

Codex

codex plugin marketplace add ora/webmcp
codex plugin add webmcp

Cursor / skills.sh

npx skills add ora/webmcp

What you get

Three skills. webmcp stops for your approval before touching any file; audit writes one report (webmcp-audit.md) and no code; verify changes nothing.

Skill What it does
webmcp The full workflow: inventory your app, propose tools per user journey, implement, verify, harden
audit Read-only agent-readiness report: scored findings across find/read/use, fix-first list, proposed tool table. Writes no code
verify Runtime check of existing tools: registration, invocation, security lint. Reports each tool as verified, failed, or could-not-verify

Plus two small MIT packages:

  • @ora-ai/webmcp-verify: a CLI that launches your installed Chrome with WebMCP enabled, lists a page's registered tools, lints them, and executes them the way an agent would. npx @ora-ai/webmcp-verify http://localhost:3000 — no flags, no Canary, no console pasting.
  • @ora-ai/webmcp-bridge: registers your existing remote MCP server's tools on document.modelContext, for the bridge strategy. Built on the official MCP SDK.

Generated page tools still need no SDK at all.

Use it

> Audit this site's agent-readiness
> Make this app WebMCP-compatible
> Verify my WebMCP tools

The plan comes first. You approve which journeys become tools, then the agent implements against the standard API: imperative registerTool tools (the default — every WebMCP runtime reads them), declarative form annotations (Chrome preview only; runtimes that read registered tools, such as ChatGPT Site tools, never see them), or a bridge to an MCP server you already run.

Principles

  • Standard API only. Generated code targets document.modelContext with feature detection and graceful no-op — no SDK required. The one library this plugin may add is @ora-ai/webmcp-bridge, only for the opt-in bridge strategy (mirroring an MCP server you already run).
  • Journeys, not endpoints. Tools map to what a visitor asks and asks for. Most sites need 3 to 10 tools. A blog needs 1.
  • Approval before implementation. Tool selection is a product decision. No code is written until you say yes; audit's only artifact is its report file.
  • Verified, not assumed. Every tool is exercised in a browser the way an agent would call it.
  • Safe by default. Server-side authorization stays mandatory. Payments and deletes stop at a reversible boundary. Honest annotations.

Requirements

  • A locally runnable website. Any stack: the skills cover React/Next, Vue, Svelte, and vanilla or server-rendered MPAs.
  • For browser verification: Chrome. npx @ora-ai/webmcp-verify launches it with the right flags itself; non-Chromium browsers need @mcp-b/webmcp-polyfill.

Development

  • node scripts/check.mjs runs the repo checks CI runs: manifest validity, version sync, skill frontmatter, link integrity.
  • npm install at the repo root sets up the git hooks (husky): conventional-commit messages via commitlint, and check + lint + tests on push. npm run lint lints the bridge package (eslint).
  • tests/fixtures/bookshop/ is a small static site for exercising the skills by hand; its README lists the expected outcome per skill.
  • evals/ holds cases for claude plugin eval (early access).
  • Issues and PRs welcome. The skills are plain markdown; the bar for changes is accuracy against the current spec and Chrome docs.

WebMCP resources

About Ora

Ora ranks how ready products are to be used by AI agents. WebMCP tools are one of the strongest signals in the usability layer. Run a free scan of your site at ora.ai, or from the terminal: npx @ora-ai/ax audit your-site.com.

Ora is a product of Era Labs; the GitHub org is ora.

License

MIT

About

Make your website agent-ready with WebMCP. Coding-agent plugin for Claude Code, Codex, and Cursor: audit, implement, and verify document.modelContext tools.

Topics

Resources

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages