Web search and fetch

A conversation can search the web and read pages, either through the search built into its model's provider or through a search service you set up in Mework. You turn web access on per conversation, choose who searches and fetches, and can filter and trim what comes back.

Turn on web access

Web access is one switch per conversation: Conversation settings → Advanced tools → Enable web search. Open conversation settings from More options → Conversation settings in the top bar, or press ⌘⇧, (CtrlShift, on Windows).

With the switch off, the model gets no web tools. With it on, it gets up to two:

• web_search whenever Search provider is not Off.
• web_fetch whenever Fetch provider is not Off; with Native, only on Anthropic Messages and AWS Bedrock models.

If the chosen provider is switched off in Settings, unavailable or missing its API key, the tool is still offered, and each call fails with a message naming the setting to change. Changing providers in Settings → Providers → Search providers never adds a tool to a conversation or removes one.

The built-in mework preset starts with the switch on, both providers on Native, Result count 5, Result compression 2,000 and Domain filter off.

A preset's settings window and each agent role have the same Advanced tools page. A role reaches the web only when the conversation that starts it does, but it can choose its own providers and domain filter, or Follow the conversation.

Search providers

Advanced tools → Search provider chooses who runs a search:

If a provider you picked is later switched off in Settings, becomes unavailable or lacks its API key, the menu reads Repair search provider, and searches fail until you fix it or choose again.

Native search by provider type

Whether Native works depends on the type of the model's provider, as shown in Settings → Providers → Model providers (see providers):

On Anthropic Messages and AWS Bedrock, the Native row opens a submenu with the tool version: web_search_20250305 (the default) or web_search_20260209, which adds dynamic filtering; for fetch, web_fetch_20250910 (the default) or web_fetch_20260209.

On a model whose type has no native search, Native cannot search, so choose a provider; ExaMCP is on out of the box and needs no API key.

Native search becomes fixed for a conversation once a search has run with it on a model that supports it, and Native fetch once a request has offered it. A fixed selector is gray, and you start a new conversation to choose again. Nothing else is ever fixed, so on a model without native search you can switch to a provider at any time. While the model's prompt cache is warm the selector is orange, because after a change the next request cannot reuse that cache.

Set up a provider

Open Settings → Providers → Search providers (⌘, opens Settings, Ctrl, on Windows). Choose a provider in the list and turn on the switch in its header.

Fetching pages

Advanced tools → Fetch provider works like the search provider, and the two are independent.

• Native: on Anthropic Messages and AWS Bedrock models, the provider fetches the page. Other provider types get no web_fetch with Native; OpenAI-style providers read pages as part of their search.
• A provider that can fetch and is switched on: fetch, Jina, Firecrawl or Querit.
• Off: no web_fetch.

A fetch provider that is later switched off, unavailable or missing its API key shows Repair fetch provider: web_fetch is still offered, and fetches fail until you fix it or choose again.

The fetch provider

The provider named fetch needs no account: Mework downloads the page itself and extracts its readable text.

• It does not run JavaScript, so a page that builds its content in the browser can come back empty.
• It reads up to 4 MiB, keeps up to 200,000 characters of text, and gives up after 30 seconds.
• It refuses pages on private or loopback addresses, except that at Full access the model may fetch a local address it names directly, such as your dev server. Redirects into your local network and links found in search results are always refused.

Domain rules

Advanced tools → Domain filter offers Use blocklist, Use allowlist and Off. Edit lists opens Domain lists, with a Blocklist and an Allowlist page: one rule per line, up to 512 rules per list.

• Only one list is in effect at a time; turning the filter off keeps both lists.
• The blocklist drops results that match a rule. The allowlist keeps only results that match, so an empty allowlist keeps nothing.
• Rules apply to search and fetch results from a provider, and fetching a blocked URL returns nothing. They do not apply to Native.

Anything else is not a rule and is ignored. A bare example.com or *.example.com is ignored, so write *://*.example.com/* to cover a domain and all its subdomains. A pattern needs a path: https://example.com is ignored, and https://example.com/ matches only the home page.

An agent role on Follow the conversation uses the conversation's filter and lists; a role with its own mode uses only its own lists.

Shaping results

Two numbers on Advanced tools limit how much of a provider's results goes into the context. Neither applies to Native, and an agent role always sets both itself.

Approvals

At Manual and Accept edits, web_search and web_fetch ask before they run; at Full access they do not (see security levels). One card covers the whole call, the query or every URL in it, and Always allow is offered. Results reach the model marked as untrusted web content.