Cloudflare Adds Web Search for Agents to AI Gateway, Partner Prices With No Markup
AI agents often guess the address of the page they need and fetch it with curl, which frequently results in a 404 error. Cloudflare has added web search to AI Gateway, its gateway for model requests. Agents now search the way a person would, through Exa, Linkup, or Ceramic.ai. Usage is paid with gateway credits at the partners' rates, with no Cloudflare markup. This makes it easier for agents to get up-to-date documentation without building their own search layer.
AI-processed from Cloudflare AI Blog; edited by Hamidun News
If you have run AI agents with internet access, you know the typical scene: the agent confidently "remembers" a page address, runs curl, and gets a 404 Not Found error. It did not search for information. It simply guessed the URL. Cloudflare says so directly and offers a fix. The company has built web search into AI Gateway, its gateway through which applications reach language models and where logs, billing, and access control are consolidated. Now an agent can search first, the way a person does in a search engine, instead of typing addresses at random. You can pay for this at the same rates as the search partners, with no Cloudflare markup.
The Web Search API launched with three providers: Ceramic.ai, Exa, and Linkup.
Why an agent needs search
A language model is trained on data up to a certain date and is "frozen" after that. Ask it about a fresh library release, a changed API, or yesterday's news, and it will answer from old knowledge or make up a plausible-sounding fabrication. Web search closes that gap: the agent sends a query, receives short structured snippets from the internet, and those snippets are inserted directly into the model's context.
Cloudflare offers its own example. An agent writing code for Cloudflare Workers may not know about new products released during their Birthday Week. With search, it can pull up current documentation and release notes instead of relying on outdated knowledge.
How it connects
There are two ways. The first is a standard REST request. If you already have a backend, a mobile app, or an external service, you send a POST to your Cloudflare account endpoint with an AI Gateway token. In the body you specify the query, the provider (for example, ceramic), and a result limit, say five.
The second way is for those building on Cloudflare Workers. There is a dedicated binding: you call env.AI.websearch with the gateway ID, the query, and the provider, for example exa, and get the results back as JSON.
Cloudflare also shows how to make search a server-side tool manually. The model, in the example Gemma 4 26B, receives a description of a web_search function and decides for itself when to call it. The Worker runs the search through env.AI.websearch, the results are returned to the model in the tool role, and the model produces the final answer. The scheme is standard tool calling; the only difference is that the entire plumbing lives on one platform.
Built-in Server Tools are promised soon. Then you will not have to describe the tools yourself: they will be part of the AI Gateway control layer, and web search will be one of the first. For now, this is an announcement, not a finished feature.
Billing, logs, and crawler rules
For teams already using AI Gateway, the main convenience is unified reporting. Search requests appear in the regular logs, are charged against the AI Gateway credit balance, and access to providers can be restricted by the same rules that apply to models. Cloudflare says it takes the partners' prices from their public price lists, with no additional markup. The BYOK (Bring-Your-Own-Key) scheme is also supported: you can connect your own provider keys. Cloudflare promises to label partners that have a Zero Data Retention policy, meaning they do not store your data, separately.
A separate point concerns crawling ethics. Partners have committed to meeting Cloudflare's requirements for "Verified bots": a crawler must identify itself honestly, respect robots.txt, and search responses must include a link to the source. For site owners, this is a way to keep control over how their content is used. For developers, it also reduces legal and reputational risk.
What to keep in mind
The announcement contains no benchmarks of result quality, no comparison of providers against each other, and no latency data, so choosing between Exa, Linkup, and Ceramic.ai will require your own tests on your queries. There is also no assessment of how much search will reduce hallucinations: that is only a logical consequence of fresh context. In addition, third-party web snippets in a prompt open the door to prompt injection, and you still need to filter them yourself.
The practical takeaway is simple. If your agent already runs through AI Gateway, connecting search takes a few lines of code, and billing and logs stay in one place. Try replacing URL guessing with search on tasks where current documentation matters, and compare the share of failed requests before and after.
- Nov 21, 2026Агенты решают, что показать юзеру: как отдать им интерфейс через MCP Apps или A2UI
- Nov 21, 2026Linger запускает 12 ИИ-моделей на вашем ПК: до 5 раз быстрее и $0 за токены
- Nov 21, 2026Sber's AutoJudge Identifies the Faulty Agent in 67% of Cases, Up from 55%
- Nov 21, 2026AWS Connects Quick AI Assistant to Seller Central and Gives Sellers 12 Months of Plus
Key Questions Answered
Why an agent needs search?
A language model is trained on data up to a certain date and is "frozen" after that. Ask it about a fresh library release, a changed API, or yesterday's news, and it will answer from old knowledge or make up a plausible-sounding fabrication. Web search closes that gap: the agent sends a query, receives short structured snippets from the internet, and those snippets are inserted directly into the model's context. Cloudflare offers its own example.
How it connects?
There are two ways. The first is a standard REST request. If you already have a backend, a mobile app, or an external service, you send a POST to your Cloudflare account endpoint with an AI Gateway token. In the body you specify the query, the provider (for example, ceramic), and a result limit, say five. The second way is for those building on Cloudflare Workers.
What are the key details on Billing, logs, and crawler rules?
For teams already using AI Gateway, the main convenience is unified reporting. Search requests appear in the regular logs, are charged against the AI Gateway credit balance, and access to providers can be restricted by the same rules that apply to models. Cloudflare says it takes the partners' prices from their public price lists, with no additional markup. The BYOK (Bring-Your-Own-Key) scheme is also supported: you can connect your own provider keys.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.