webmcp — agent skill

Detect whether a website exposes WebMCP (Web Model Context / navigator.modelContext) browser tools, and call them correctly — instead of scraping the DOM and guessing which button does what.

WebMCP lets a site register typed, agent-callable tools (navigator.modelContext.registerTool({ name, description, inputSchema, execute })). This skill teaches your coding agent to find those tools on any page, read their schemas, and invoke them safely.

Install

npx skills add Array-Ventures/webmcp

This installs the skill into your agent's config (Claude Code, Cursor, and other SKILL.md-compatible agents). List or remove it later with npx skills ls / npx skills rm webmcp.

Prefer to install by hand? Copy SKILL.md, references/, and scripts/ into ~/.claude/skills/webmcp/ (global) or .claude/skills/webmcp/ (per project).

What it does

• Detect — one probe reports whether navigator.modelContext / navigator.modelContextTesting are present, whether the page is a secure context, and the full list of registered tools with their JSON-Schema inputs. It also catches declarative tools defined via HTML toolname attributes that the JS API can miss.
• Call — invokes a tool with the correct argument encoding (the testing API wants a JSON string, the single most common mistake), and reads back the structured result.
• Diagnose — when a WebMCP page lists zero tools, it tells you which of the real causes it is: the testing flag is off, the page isn't a secure context, tools register late or behind a route/login, or the site's own registration crashed (e.g. a non-HTTPS origin in an exposedTo list throwing a SecurityError that aborts every registration).
• Stay safe — runs read-only tools freely, but pauses for confirmation before anything that looks like it writes, submits, books, pays, or authenticates.

Requirements

WebMCP is experimental and only works in a WebMCP-capable browser:

1. Chrome Canary/Beta (or recent Chromium).
2. Enable chrome://flags/#enable-webmcp-testing and relaunch.
3. The target page must be HTTPS or localhost.

The skill drives the browser through your agent's Playwright/MCP browser when one is connected. If not, the bundled scripts/webmcp-probe.mjs launches its own Chromium:

# Inspect the page's WebMCP surface (prints one JSON report)
node scripts/webmcp-probe.mjs https://render.com

# Inspect, then invoke one tool
node scripts/webmcp-probe.mjs https://render.com --call render.docs.search --args '{"query":"background worker"}'

When the bundled Chromium lacks the flag, point at Canary/Beta:

WEBMCP_CHROME="/Applications/Google Chrome Canary.app/Contents/MacOS/Google Chrome Canary" \
  node scripts/webmcp-probe.mjs https://render.com --headed

Try it

https://render.com ships real WebMCP tools (docs search, docs-by-slug, blog/articles indexes). Ask your agent: "does render.com have webmcp tools? list them, then search the docs for 'background worker'."

Layout

SKILL.md                       # the skill: detect → list → inspect → execute, safety, diagnosis
references/webmcp-internals.md  # API surface, schemas, gotchas, troubleshooting tables
scripts/webmcp-probe.mjs        # standalone fallback runtime (own Chromium; no third-party calls)
evals/evals.json                # test prompts used to validate the skill

Privacy

The skill and script only ever contact the URL you give them. No directory service, telemetry, or analytics. (Installing via npx skills reports anonymous install counts to skills.sh — that is the CLI's behavior, not this skill's.)

License

MIT © Array Ventures