webagents.md
https://github.com/browser-use/webagents.md- Category
- AI Agents
- Rank
- No. 1690Tools index
Previous survey · No. 1697 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- browser-use
- GitHub
- 12 stars
- Date
About
Proposal letting websites expose tools for AI agents to call directly in the browser. By browser-use.
What it does
It defines a browser-side contract between websites and AI agents. Sites describe available operations in Markdown and advertise that description through page metadata. The Python SDK discovers and parses the manifest, gives an LLM typed TypeScript context, then runs the model’s generated JavaScript inside the active browser session.
Why it's ranked here
The idea is unusually coherent: website owners publish the supported automation surface, while agents compose several operations with ordinary code. The repository backs the proposal with parsing, discovery, type generation, validation, browser execution, tests, and a working bookstore demo. Still, its usefulness depends on website adoption, and the specification remains explicitly proposed rather than established.
What's good
The manifest doubles as human documentation and model context, preserving operational notes such as rate limits and workflows. Typed parameters, optional defaults, and declared outputs become TypeScript guidance for code generation. Two Markdown syntaxes support hand-written documentation and compact configuration. Website builders also get construction, serialization, and validation helpers, while agent builders get discovery and execution plumbing.
Tradeoffs
Authentication is outside the specification, so operations inherit the browser’s existing cookies, tokens, and permissions. The SDK executes model-written JavaScript in that privileged context, making site trust and runtime controls important. Missing output declarations fall back to an untyped result. Validation returns warnings for a small set of structural problems rather than enforcing a complete security or compatibility contract. The caller must supply the model loop and browser page.
How to use it well
It best suits website teams willing to publish a deliberate automation surface, and agent teams experimenting with code-generating browser workflows. Start with narrow, well-described operations, explicit result types, limits, and failure behavior. Keep browser permissions constrained and inspect generated execution. It does not replace authentication design, backend APIs, MCP integrations, or a production agent orchestration layer.
Technical notes+
pyproject.toml packages src/webagent with Hatchling, requires Python 3.11+, and declares Pydantic and HTTPX dependencies. src/webagent/discovery.py locates manifests with regular expressions over HTML metadata, while src/webagent/parser.py supports heading and compact syntaxes. src/webagent/codegen.py emits a declare const global TypeScript block. src/webagent/client.py combines raw Markdown with declarations, exposes an OpenAI-style execute_js schema, wraps code in an async browser expression, aliases global to globalThis, and converts exceptions to JSON errors. A notable parser mismatch exists: every level-two heading becomes a Tool, despite README.md saying non-tool sections are preserved as context. tests/test_webagent.py covers parsing, generation, serialization round trips, and raw-content preservation, but not discovery or browser execution.
Observed
- License
- MIT License
- Primary language
- Python
- Installation
- Published package surface documented as pip install webagents-md
- Runtime requirement
- Python 3.11 or newer
- Interface
- Python library for manifest discovery, parsing, generation, validation, serialization, and browser execution
- Core dependencies
- Pydantic 2 or newer and HTTPX 0.25 or newer
- Browser model
- Executes generated JavaScript through a caller-provided page object with an evaluation method
- Tests
- Repository includes core tests for both manifest formats, TypeScript generation, serialization round trips, and content preservation
Read from README.md, pyproject.toml, src/webagent/site.py, src/webagent/types.py, src/webagent/client.py, src/webagent/parser.py, src/webagent/codegen.py, src/webagent/__init__.py, src/webagent/discovery.py, src/webagent/serializer.py, demo/demo.py, demo/run_agent.py, tests/test_webagent.py, LICENSE, demo/README.md.
What it can do
Enable websites to expose callable tools for AI agents
Website with defined tool interfaces → AI-accessible tool endpoints
Allow AI agents to directly interact with web applications
AI agent requests and web application tools → Direct tool execution results
Facilitate browser-based AI agent automation
AI agent commands and browser context → Automated web interactions
Provide standardized interface for AI-web integration
Web application functions and AI agent calls → Standardized communication protocol
Execute web-based tools through AI agent requests
Tool parameters and AI agent instructions → Tool execution responses
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.