Web Tools

Glue ships three web tools by default: web_search, web_fetch, and web_browser. The surface is a little wider than in most coding agents because I use Glue for scraping and automation about as much as for coding — nothing hype, just more practical capability than you'd typically find.

Everything below works out of the box. Providers with API keys light up automatically, and DuckDuckGo gives web_search a zero-config fallback. Browser backends swap with one config line.

The tools at a glance

Tool	What it does	Status
web_search	Query Brave, DuckDuckGo, Tavily, or Firecrawl. Auto-detects keyed providers with DuckDuckGo fallback.	shipping
web_fetch	HTML → cleaned markdown, PDF → text. OCR and Jina fallbacks.	shipping
web_browser	Drive a real Chrome over CDP. Six actions, seven backends.	experimental

Search

shipping

Four providers, one interface. Glue prefers an explicitly configured provider. Otherwise it tries the first configured provider it finds and falls back to DuckDuckGo, which needs no API key.

Provider	Strength	Env var
DuckDuckGo	Zero-config fallback via the HTML endpoint. No account, no key.	None
Brave	General results, straightforward developer plan.	BRAVE_API_KEY
Tavily	Agent-focused; returns short summaries with results.	TAVILY_API_KEY
Firecrawl	Search plus a crawl/scrape API on the same key.	FIRECRAWL_API_KEY

~/.glue/config.yaml — pin a search provider

web:
  search:
    provider: brave # brave | tavily | firecrawl | duckduckgo

Fetch

shipping

web_fetch reads a URL and returns it in a form the model can use — HTML cleaned to markdown, PDFs as extracted text. It has two fallbacks for the awkward cases:

• OCR for scanned PDFs. If a PDF has no extractable text layer, Glue passes the pages to a vision model (Mistral or OpenAI) and returns the OCR'd text instead. Useful for government documents, old filings, or anything scanned from paper.
• Jina Reader for pages that extract poorly with the default cleaner. Set JINA_API_KEY and Glue routes difficult fetches through Jina.

~/.glue/config.yaml — fetch with OCR + Jina

web:
  fetch:
    jina_api_key: ${JINA_API_KEY} # optional, for hostile pages
  pdf:
    ocr_provider: mistral # mistral | openai
    mistral_api_key: ${MISTRAL_API_KEY}

Browser

experimental

web_browser drives a real Chrome over the DevTools Protocol. The CDP backend is experimental — it works across all providers but is newer than the rest of Glue, and rough edges should be expected.

Actions

Six verbs, all running against a single browser session that persists across tool calls within the same turn. The agent can navigate, fill a form, submit, and read the result without reopening the tab.

Action	Use for
navigate	Open a URL. Waits for the page load event by default; pass wait_until to switch (domcontentloaded | networkalmostidle | networkidle).
click	Click any selector in the live DOM.
type	Fill an input.
screenshot	Full page or a single element, saved to disk.
extract_text	Cleaned markdown of the current page (capped at ~50k tokens).
evaluate	Run arbitrary JavaScript in page context and return the result.

Backends

All seven implement the same BrowserEndpointProvider interface, so swapping is a one-line config change rather than a code change.

Backend	Runs on	Best for	Status
local	Your machine (Puppeteer + Chrome)	Iteration. headed: true lets you watch it work.	shipping
docker	Local browserless/chrome container	Keeping the browser off your host. Container dies with the session.	shipping
browserbase	Cloud — browserbase.com	Hosted sessions with replays.	experimental
browserless	Cloud or self-hosted — browserless.io	Scale and self-hosting.	experimental
steel	Cloud — steel.dev	Agent-focused cloud sessions.	experimental
anchor	Cloud — anchorbrowser.io	Agentic browser sessions with live view.	experimental
hyperbrowser	Cloud — hyperbrowser.ai	Agent-focused cloud sessions with fast launches.	experimental

~/.glue/config.yaml — switch browser backends

web:
  browser:
    backend: docker # local | docker | browserbase | browserless | steel | anchor | hyperbrowser
    docker:
      image: browserless/chrome:latest
      port: 3000

Cloud backends need an API key. Prefer env vars or ~/.glue/credentials.json over committing keys into config.yaml.

Credentials

Every API key resolves in the same order: environment first, then ~/.glue/credentials.json, then — if you really insist — config.yaml. Nothing auto-reads project-local config.

What needs a key	Env var
Brave search	BRAVE_API_KEY
Tavily search	TAVILY_API_KEY
Firecrawl search	FIRECRAWL_API_KEY
Jina fetch fallback	JINA_API_KEY
Mistral OCR	MISTRAL_API_KEY
OpenAI OCR	OPENAI_API_KEY
Browserbase browser	BROWSERBASE_API_KEY + BROWSERBASE_PROJECT_ID
Browserless browser	BROWSERLESS_API_KEY
Steel browser	STEEL_API_KEY
Anchor browser	ANCHOR_API_KEY
Hyperbrowser browser	HYPERBROWSER_API_KEY

Pairing with runtimes

Web tools compose with Glue's runtime — pair them when the work is risky or you don't want the traffic on your host.

Goal	Suggested pairing
Quick fetch or a search	Host runtime + whatever provider you have.
Clicking through an untrusted site	Docker runtime + Docker browser.
Reading a scanned filing	Any runtime + fetch with OCR.
Scale scraping	planned Cloud runtime + cloud browser.

Full matrix of runtime × browser combinations lives on the Runtimes page.

Web tools guide → · Browser automation guide →