Web

web_fetch downloads a page and returns its main content as Markdown:

Scripts, navigation, headers and footers, and decorative images are stripped, and links are resolved against the page URL.
Text and JSON responses are returned as-is.
Long pages come back in chunks (start / max_length).
It asks per domain. "Always allow" saves a rule such as web_fetch(domain:go.dev), which also covers subdomains.
A redirect to a different host is reported rather than followed, so it can't sidestep a domain rule.
Link-local and cloud-metadata addresses (169.254.169.254 and similar) are blocked.
Responses are capped at 5 MB and 30 s, and cached for 15 minutes.

web_search appears when a search backend is configured. Larik detects one from the environment:

Backend	Setting
Brave Search	`BRAVE_API_KEY`
Tavily	`TAVILY_API_KEY`
SearXNG (self-hosted; enable the JSON format)	`SEARXNG_URL`

You can also set it explicitly:

json

{ "web": { "search": { "provider": "brave", "api_key_env": "MY_BRAVE_KEY" } } }

Trust and permissions:

Search settings are honored only from personal files, since a shared .larik/settings.json could otherwise send your queries to its own server. Shared files can only disable web tools ("web": {"fetch_disabled": true} or {"search": {"disabled": true}}).
Both tools ask before running (web_search can be always-allowed).
Plan mode asks for them rather than blocking them, because research is part of planning.
Web content is marked as untrusted data for the model.

Browser

For pages that need JavaScript, a sign-in or clicking through, Larik can drive a real Chrome window. It's off by default; turn it on in your personal config:

json

{ "browser": { "enabled": true } }

Chrome (or Chromium) must be installed; set "chrome_path" if Larik can't find it, and "headless": true to hide the window. Chrome starts on the first browser call and keeps its own profile under ~/.local/share/larik/browser-profile, so sign-ins you make in that window last between sessions.

The model gets these tools:

Tool	Does
`browser_navigate`	Opens a URL and returns a snapshot of the page
`browser_snapshot`	The page's text, plus its links, buttons and fields, each with a ref such as `e12`
`browser_screenshot`	An image of the viewport, the full page or one element; `labels` draws each ref on it
`browser_click`	Clicks (or hovers) an element by ref with a real mouse event
`browser_type`	Types into a field by ref, key by key; optionally presses Enter
`browser_select`	Picks options in a `<select>`
`browser_press_key`	Presses Enter, Escape, Tab, arrow keys, PageDown…
`browser_history`	Back and forward
`browser_tabs`	Lists, opens, selects and closes tabs; tabs a page opens become active
`browser_eval`	Runs a JavaScript expression in the page
`browser_console`	Console messages, exceptions and dialogs since the last call

browser_navigate asks per domain like web_fetch ("always allow" saves browser_navigate(domain:github.com)), only opens http(s), and refuses link-local and cloud-metadata addresses. Clicks and typing ask per call unless you allow the tool.
Plan mode asks before opening a page and blocks everything that could change something on a site. browser_snapshot, browser_screenshot and browser_console only read, so they never ask.
Screenshots need a model that can see images. They're JPEGs at one pixel per CSS pixel (a 1280×900 viewport is about 1,500 tokens on Claude) and stay in the conversation until it's compacted.
Only personal files can enable the browser or choose its binary; a shared .larik/settings.json can only switch it off.
Alerts are accepted and confirm or prompt dialogs dismissed automatically, and noted in browser_console.

Generated from README.md · section “Web”. Edit that file to change this page.

Web

#Browser

Browser