--- title: "让助手搜索网络 | Busymate AI" description: "按工作区开启网络搜索和网页阅读，选择网站、设置限额，并测试助手实际会得到的结果。" last_updated: "2026-09-23T16:47:50+03:00" --- # 让助手搜索网络 | Busymate AI Source: https://busymate.ai/zh-Hans/docs/guides/web-access Last modified: 2026-09-23T16:47:50+03:00 Busymate AI lets your mate look things up on the web when your own knowledge does not have the answer — today's news, a price, a product's specification — by **searching the web** and **reading a page**. It always tries your knowledge first, says which it used, and links the page when it answers from the web. Everything here lives in **Console → Web access**. Every control saves itself and is live the moment the page says **Saved**; **Undo** writes the previous values back. There is no draft and nothing to publish. ## 1. Switch it on Web access is **off** for a new workspace: nothing reaches the web until an admin turns on **Web access**. Under it, **Search the web** and **Read a page** can each be turned off on their own — for example, read the pages your customers paste in, but never search. The page opens with one plain sentence about what your mate can do **right now**, read from the saved settings rather than from anything still being typed, and where those settings come from. A workspace that switched on live web search in an earlier version of its settings starts with **Search the web** on and **Read a page** off, and the page says so; the first change you make here takes over from that older setting. ## 2. Choose the sites - **Only these sites** — when the list is empty, any public site may be used. When it has entries, only those sites and their subdomains are searched or read. - **Never these sites** — always refused, even when the same site is also allowed. Enter a site the way you would say it: `example.com` or `*.example.com`. A full address such as `https://example.com/page` is refused with a note, because the lists match on the site, not on a path. Both lists apply to searching and to reading a page — including every address a link redirects to, so a short link cannot lead your mate to a site you blocked. A site you block inside one you allow (allow `example.com`, block `forum.example.com`) is also left out of search results. ## 3. Set the limits Searches and page reads count together. | Limit | What it bounds | Range | |---|---|---| | Per conversation | Lookups one conversation may make | 1 – 50 | | Per day | Lookups the whole workspace may make in a UTC day | 1 – 100,000 | A lookup counts when it is **attempted**, so a site that keeps failing is not a free retry loop. When a limit is reached, your mate tells the person plainly that it cannot look anything else up — it never pretends it searched. ## 4. Decide how it answers - **Cite sources** (on by default) — your mate names and links the page it used, never cites a page it did not open, and never invents an address. - **Safe search** (on by default) — asks the search to leave out explicit results. ## 5. Try it The **Try it** box at the bottom of the page runs one real search or one real page read under the **saved** settings and shows exactly what your mate would get: the answer with its sources, or the reason it would refuse. Tests do not count toward your limits and are not recorded as usage; they have a small allowance of their own, six every ten minutes. ## What is never read Page reading refuses, before a single byte is fetched: addresses that are not `http` or `https`, addresses that carry a user name or password, addresses that name a port, and anything that is not on the public internet — local, private and cloud-metadata addresses, including a public name that resolves to a private one and a redirect that lands on one. A site that asks assistants not to read a page (`robots.txt`) is respected. The text of a page is bounded, and your mate is told when it was cut. ## Manage from any MCP client The same settings are available to agents and scripts: `get_tenant_web_access`, `set_tenant_web_access` and `test_tenant_web_access` over MCP, or `GET /api/v1/web-access`, `POST /api/v1/web-access` and `POST /api/v1/web-access/test` over the [REST API](https://busymate.ai/zh-Hans/docs/guides/api). A change names only the fields to change — every other field is kept — and is live when it returns. The Insights metrics `web_searches` and `web_fetches` count lookups per day. Each lookup is also recorded in the events trail with its site and outcome, never the question or the page text. ## Verify 1. In **Try it**, search for something your knowledge does not cover — the answer arrives with its sources. 2. Add that answer's site to **Never these sites** and run the same test — it is refused with the blocked-site reason. 3. Turn **Web access** off and ask your mate the question in the chat — it says it cannot look this up here and answers from your knowledge. ### Does my mate search the web for every question? No. It answers from your own knowledge first and goes to the web only when that knowledge does not have the answer, or when the question is about something that changes, such as news or prices. It says which one it used. ### Does it cost me more when it searches? Each lookup is bounded by the per-conversation and daily limits you set, so you decide the ceiling. Tests from the Try-it box never count toward them. ### Can it read pages behind a login or on my private network? No. It only reads public addresses. Private, local and cloud-metadata addresses are refused before anything is fetched, and so is any address carrying a user name or password. ### Can I allow only my own sites? Yes. Add them to **Only these sites** and both searching and page reading stay on those sites and their subdomains. ## Next - **[Teach your assistant your own content](https://busymate.ai/zh-Hans/docs/guides/knowledge)** — the knowledge your mate tries first. - **[REST API](https://busymate.ai/zh-Hans/docs/guides/api)** — keys, scopes and the rest of the workspace API.