OAI

[OAI @oai_official]

Hey everyone,
We've been talking a lot about building powerful AI agents with MCP (for tools & context) and A2A (for agent coordination).

But there's still a big practical gap: Most agents are crawling the public web completely blind.

They make 40+ requests per site just to figure out what the website can actually do — actions, policies, rate limits, etc. This slows everything down and causes a lot of brittle scraping.

To solve this, I built W2A (Web-to-Agent) Protocol — a lightweight, web-native third layer. It adds one simple file: agents.json (usually at /.well-known/agents.json)

This file declares what the website offers to agents in a standardized way. Agents can read it in one request instead of blind crawling.
✅ Super low friction:
-> Auto-generates from your existing https://t.co/LNoq47keiZ, Open Graph, and sitemap signals
-> Gives a confidence score
-> One-file deploy (works with Vercel, Next.js, WordPress, Shopify, etc.)

Just tried it on https://t.co/pxNL0QgL1x and got 92% confidence in under 10 seconds.

Try the free generator here:
https://t.co/lKvWr6KrHa
It complements MCP + A2A perfectly:
MCP → agents ↔ tools
A2A → agents ↔ agents
W2A → websites ↔ agents {Missing Layer}

Would love your feedback as RAG / agent builders:
- How painful is web interaction for your agents today?
- Would a standard agents.json help you?
- Any features you'd want to see in the spec?

Happy to answer questions or review any agents.json you generate.

Looking forward to your thoughts!

[OAI @oai_official]

langchain-w2a v0.2.0 is live on PyPI with full LangChain
standard test compliance.

✓ ToolsUnitTests — 12 passed, 0 warnings
✓ ToolsIntegrationTests — pass against live site
✓ Docs PR open: langchain-ai/docs#4021

pip install langchain-w2a

from langchain_w2a import W2AToolkit
toolkit = await W2AToolkit.from_url("https://t.co/uzfOR4780U")

#LangChain #AIAgents #OpenSource

[OAI @oai_official]

Full stack is now live:
Website owners
→ 2-minute manifest at https://t.co/wyRaJnPXta

Python devs → pip install w2a
JS/TS devs → npm install w2a-client
Claude/Cursor users → npx w2a-mcp
LangChain users → pip install langchain-w2a (just announced)

All open source (Apache 2.0).
Which part are you going to try first?

#AIAgents #W2AThe

[OAI @oai_official]

14 requests. 7.5 seconds. Still uncertain.
vs.
1 request. 2.9 seconds. Done.

Same site. Same agent. Same task.

The only difference: one JSON file.

This is what /.well-known/agents.json does for AI agents visiting your website.

Try it on any site
→ https://t.co/ArWizxgewG

@cursor_ai
#AIAgents #MCP #A2A #OpenSource

[OAI @oai_official]

For agent developers:

Python: pip install w2a

JavaScript / TypeScript: npm install w2a-client

Claude / Cursor users: npx w2a-mcp

One command and your agent can understand any W2A-enabled website.

#AIAgents #W2A

[OAI @oai_official]

robots.txt solved the 1994 crawling problem. sitemap.xml solved the 2005 discovery problem. agents.json solves the 2026 agent problem.

Apache 2.0. Open standard.
Same governance model as @A2AProtocol under @linuxfoundation.

Spec: https://t.co/y0BoqImEsz

@IETF @w3c @TheOrderAI

[OAI @oai_official]

For website owners seeing bot traffic spike in your logs:

One file at /.well-known/agents.json tells agents exactly what you offer.

Well-behaved agents read it first.
Blind crawling stops.

Generate yours free: https://t.co/wyRaJnPXta

@Shopify @stripe @vercel @netlify @WordPress @ghost

[OAI @oai_official]

Everything live for W2A - Web2Agent Protocol:

For website owners:
→ https://t.co/wyRaJnQviI (no code)

For Python developers:
→ pip install w2a
→ pip install langchain-w2a

For JS/TS developers:
→ npm install w2a-client

For Claude / Cursor users:
→ npx w2a-mcp --url https://t.co/uzfOR47FQs

One standard. Every runtime covered. https://t.co/y0BoqInci7

#AIAgents #MCP #LangChain #OpenSource

[OAI @oai_official]

For AI Agents Devs

Shipped W2A - an open standard for agent-readable websites.

The problem: agents crawl websites blind. 40-50 requests just to understand what a site does.

The fix: /.well-known/agents.json declares your site's
skills, actions, inputs, and policies. One read, done.

Works with:
- pip install w2a (Python)
- npm install w2a-client (JS/TS)
- npx w2a-mcp (Claude Desktop / Cursor)
- pip install langchain-w2a (LangChain)

A2A compatible. Apache 2.0.

Live tools: https://t.co/wyRaJnPXta
GitHub: https://t.co/y0BoqImEsz

Feedback welcome, especially from anyone building
agents that interact with external websites.

[OAI @oai_official]

Built something that solves a problem I kept hitting with RAG pipelines - agents don't know what a website can do before crawling it.

W2A adds /.well-known/agents.json to any site. One file declares all endpoints, inputs, outputs, and auth.

Your RAG agent reads it once instead of crawling multiple pages/URLs.

Check any site in one command:
npx w2a-mcp@latest --url https://t.co/uzfOR4780U --check
Python SDK: pip install w2a

Full spec: https://t.co/xXOsWg2VAj

Would love feedback from anyone building RAG pipelines.

[OAI @oai_official]

The W2A MCP server ships with two CLI commands:

npx w2a-mcp@latest --url https://t.co/xXOsWg2VAj --check
npx w2a-mcp@latest --url https://t.co/xXOsWg2VAj --list

Then add to Claude Desktop config and every skill becomes a tool Claude can call.
https://t.co/09tBSEXVWh https://t.co/xXOsWg2VAj
#MCP #Claude #AIAgents #OpenSource

[OAI @oai_official]

Claude / Cursor users — this one’s for you.

Add this single block to your Claude Desktop config:

"mcpServers": {
"w2a": {
"command": "npx",
"args": ["w2a-mcp"]
}
}

After that, every site with a W2A manifest instantly becomes a usable tool.

No more vague “browse this website” prompts.
Just npx w2a-mcp and go.

Tried it with https://t.co/pxNL0QgL1x yesterday — worked immediately.

Who else is using this?

#AIAgents #W2A #A2A

[OAI @oai_official]

cc @cursor_ai @claudedesktop

[OAI @oai_official]

The agent protocol stack is now complete:
MCP = agents ↔ tools
A2A = agents ↔ agents
W2A = websites ↔ agents

Website owners: 2-minute manifest
Agent devs: one import
Claude users: one npx command

All live today.

#AIAgents #W2A

[OAI @oai_official]

For LangChain users:
from langchain_w2a import W2AToolkit
toolkit = W2AToolkit.from_url("https://t.co/tblFV3tFNN")

One toolkit = all declared skills as typed LangChain tools.
pip install langchain-w2a

For JS/TS devs:
npm install w2a-client

#AIAgents #W2A

[OAI @oai_official]

Tired of agents crawling your site blindly with 40+ requests?

Go to https://t.co/Yma3VotUTN
→ Enter your URL
→ Download agents.json
→ Drop it in /.well-known/agents.json

Takes <2 minutes. Your site now has a proper voice.

#AIAgents #W2A

[OAI @oai_official]

Before: Writing custom scrapers + handling 40+ blind requests per site.
After:
from w2a import discover
skills = await discover("https://t.co/4as3udMK4i")

One line → structured skills with auth, policies and rate limits.
pip install w2a

#AIAgents #W2A #A2A

[OAI @oai_official]

Nailed it.

We built exactly that missing layer.
W2A (Web-to-Agent) Protocol gives any website a standardized voice for external agents.
One simple file at /.well-known/agents.json declares skills, actions, auth, policies, and rate limits.External agents can read it before they start crawling — no more blind guessing with 40+ requests per site.

It’s the outside-in approach:
→ Rover solves inside-out (embedded agent for your own users)
→ W2A solves outside-in (external agents discovering what your site can do)

The two are highly complementary. A Rover-enabled site with a W2A manifest becomes fully agent-ready from both directions.

Your paper maps 5 agent-web architectures. W2A is essentially the open, lightweight standard for the missing one.

Would love your thoughts on how it fits (especially with DOM-based or hybrid architectures).

Live today, Apache 2.0, and you can try the auto-generator here: https://t.co/lKvWr6KrHa