Web tools
For AI Agents: the Markdown version of this page is at https://ankole.agentbull.com/en-US/docs/web-tools/index.md. The documentation index is at https://ankole.agentbull.com/en-US/llms.txt.
Agents use web_search to discover public web pages and web_fetch to read specific URLs. With Provider-hosted web search enabled, Ankole uses the current model Provider’s native search when the model supports it. To use a separate search Provider, disable Provider-hosted web_search and bind the web_search profile. web_fetch falls back to the Worker’s built-in rendered page reader when no web_fetch Provider is configured.
Select the correct tool
| Capability | Use it for | Execution surface |
|---|---|---|
web_search |
Finding public pages by query, recency, or domain scope | Main Agent and Background Agent Jobs |
web_fetch |
Extracting readable text from one or more known public URLs | Main Agent and Background Agent Jobs |
| Browser automation | Signing in, clicking, typing, paginating, capturing screenshots, and driving interactive web apps | Background Agent Jobs only |
Browser automation is provided by the browser Skill (ankole-runtime: background_job), so it never appears in the web_search or web_fetch Provider dropdowns and cannot run inline during a main-Agent chat turn. Whenever a task requires interactive browsing, ask the Main Agent to delegate the work to a Background Agent Job, which runs asynchronously without blocking the chat and reports back with questions, progress, or final results.
Configure Web tools
Add Providers
- Open Console → LLM Providers.
- Choose a Provider kind that supports
web_search,web_fetch, or both. - Enter the API key along with any provider-specific fields.
- Save the Provider and confirm that it is enabled.
You can configure multiple instances of the same Provider kind—for example, two Bright Data SERP instances targeting different countries or serving separate Agents.
Assign separate Providers to an Agent
Disable Provider-hosted web_search for the target Agent or its inherited default before using the separate search profile. Binding or clearing a profile does not change that switch.
- Open Console → Agents and select the Agent.
- Under Model profiles, locate
web_searchand choose a search-capable Provider. - Locate
web_fetchand choose a fetch-capable Provider (or leave it empty to use the Worker fallback). - Save both profiles and start a new conversation.
Because web_search and web_fetch are Provider-only profiles, you select only a Provider instance—no model name or context window is needed—and the Console automatically filters each dropdown to Providers that declare that capability. If a dropdown is empty, add a compatible Provider under LLM Providers first; see Quick start.
Current built-in Providers
Control Plane Plugins can register additional Provider kinds; the following web Providers ship built into Ankole.
web_search Providers
| Provider | Also supports web_fetch |
Main difference | Use it when |
|---|---|---|---|
| Parallel | Yes | Combines search and extraction in one Provider; accepts a research objective, multiple queries, modes, and a character budget | You want a single credential for search and page reading, or run research-heavy queries |
| Bright Data SERP | No | Backed by a SERP API; requires a Zone and supports country, language, and Google domain targeting | You need strict control over search geography, language, or localized SERPs |
| Jina Search | No | Exposes region, location, language, pagination, cache, and search-engine parameters | You want direct web search with fine-grained region, pagination, or cache controls |
| AgentBull Cloud | Yes | Unifies search and page extraction with model capabilities (llm, embedding, rerank, image_generate), supporting source scopes, time ranges, and cache bypass |
You prefer a single API key across models and web tools, or need source and recency filtering |
| Cloudflare | Yes | Runs Cloudflare Web Search through the account’s AI Gateway; lets you pick the search engine (ceramic, exa, or linkup) and returns at most 10 results per search |
You already bill AI usage through Cloudflare AI Gateway |
You can bind a single Parallel, AgentBull Cloud, or Cloudflare Provider instance to both web_search and web_fetch. By contrast, Jina Search and Jina Reader are distinct Provider kinds—create one instance of each and bind them to their respective profiles, even when both share the same Jina API key.
web_fetch Providers
| Provider | Also supports web_search |
Main difference | Use it when |
|---|---|---|---|
| Parallel | Yes | Reuses your Parallel Search Provider and API key for page extraction | You already use Parallel Search and want a single Provider entry |
| AgentBull Cloud | Yes | Reuses your AgentBull Cloud Provider and API key for public page extraction | You already use AgentBull Cloud for models or search |
| Jina Reader | No | Converts public web pages to clean Markdown with link retention, CSS target/wait selectors, caching, engine choice, and token caps | You want clean article Markdown or extraction scoped to specific CSS selectors |
| Cloudflare | Yes | Converts public web pages to Markdown through Cloudflare Browser Run using the lightweight kitesurf engine by default (with optional chromium rendering, wait selectors, navigation options, and request-rejection patterns) |
You already use Cloudflare for models or search and want fast edge Markdown extraction |
web_fetch reads public HTTPS pages as text—it cannot authenticate behind login walls and is not meant for downloading binary files such as PDFs, images, archives, audio, or video.
When web_fetch is left unbound, Ankole uses the Worker’s built-in rendered fallback (which does not appear in the Console Provider list). That fallback renders the page to extract visible text, but cannot click elements, type into inputs, persist login sessions, or capture screenshots, nor does it grant interactive browser automation to the Main Agent. Similarly, CSS wait and target selectors on a web_fetch Provider help capture late-rendered DOM text, not multi-step interaction—use browser automation in a Background Agent Job whenever a site requires signing in, clicking, or filling forms.
Ask the Agent to use Web tools
No special slash command is required—simply describe what you want researched or read:
Find three relevant announcements from this week. Read the original pages, compare the changes, and include source links.
The Agent runs web_search first and then calls web_fetch on the most relevant URLs; if you already have a URL, paste it directly and ask for a summary or extraction.
Ask the Agent to use browser automation in a Background Agent Job instead whenever:
- the page sits behind a login;
- the Agent must click tabs, paginate, or submit forms before the data appears;
- you need a visual screenshot or layout inspection of the rendered page; or
- content loads only after multi-step client-side interaction.
For straightforward search, public documentation reading, and multi-source synthesis, web_search and web_fetch are faster and lighter.
If a Web task does not work
Search or fetch failed
Verify these points in order:
- Check the selected search path: Provider-hosted search requires native support in the current model; with that switch off,
web_searchrequires an enabled search Provider.web_fetchfalls back to the Worker reader when unbound. - Confirm that the chosen Provider kind supports that capability.
- Verify that the Provider is enabled and its API key and required settings (such as a Bright Data Zone) are valid.
- Start a fresh conversation if you updated the Agent’s profiles after the current chat began.
- Confirm that the target URL is a public HTTPS page that does not require authentication.
Browser automation did not run
Confirm that the browser Skill is enabled for the Agent and that the Agent can launch Background Agent Jobs—never try to bind browser automation inside the web_search or web_fetch profile slots. If the task is already running inside a Background Agent Job, check Worker health under Console → Workers and inspect any browser or network error reported in the Job log.