Webagents.Md

webagents.md

https://github.com/user-attachments/assets/27542a07-350f-4c26-8fb4-36d8c8fc7e1a

This is a proposed spec and Python SDK that lets websites expose tools for AI agents to call directly in the browser. A site publishes a Markdown file describing its tools, and any agent can discover and use them.

If adopted broadly, `webagents.md` is how AI agents navigate the web autonomously--not by clicking through interfaces built for humans, but by calling functions that websites explicitly provide to them.

**How it works:** A website publishes a `webagents.md` file listing its tools and adds a `` tag for discovery. The SDK detects the tag, parses the manifest, and converts the tools into TypeScript declarations. The LLM gets those declarations as context plus a single `execute_js` tool.

The agent writes code like `await global.searchProducts("red shoes")`, and the runtime executes it in the browser via Playwright. Multiple calls can be chained in one shot.

**What the SDK provides:**

• For builders of AI agents, it detects a site's manifest, parses the tools, generates TypeScript declarations for the LLM, and executes LLM-written code in the browser.
• For website developers, it lets them build and validate manifests programmatically.

Why `webagents.md` exists

Most web agents today behave in one of two ways:

• They view websites and simulate clicks and keystrokes, pretending to be a human. This is not deterministic and can be time-consuming, token-intensive, and prone to failure.
• They rely on backend-only integrations (MCP servers, custom APIs, OpenAPI specs) that are fragmented and require configuration.

At the same time, we've learned a few key things about LLMs:

• LLMs are much better at writing code than at making tool calls. They've seen millions of TypeScript APIs in training, but only a smaller set of contrived tool-calling examples.
• When an LLM can write code, it can chain multiple c