Web Extract
Steinkauz supports Web Extract as a Native Tool inside chat. When the tool is enabled and at least one Firecrawl Tool backend is eligible for the current chat provider, the model can read a single page URL and use the extracted text in its reply.
This is not Web Search (a query returns hits). It is also not computer-use: the model does not click, type, or log in. Steinkauz never opens the website itself.
Tool backends
Workspace admins configure an ordered list of Firecrawl backends under Settings → Native Tools. The first backend that is allowed for the turn is used (no automatic retry on another backend if a call fails).
| Backend | Who can use it | Environment class | Billing |
|---|---|---|---|
| Firecrawl (one entry per configured instance) | All plans; usable on Gateway and BYOK chat when the instance’s Environment class is high enough for the chat provider | Inherited from that Firecrawl provider connection | You pay Firecrawl Cloud, or you run OSS; Steinkauz does not add an extract fee (invocations are audited) |
Example: order Firecrawl Cloud (typically E1) first, then a private Firecrawl instance at a higher Environment class. Gateway or Cloud-backed chats use Cloud when eligible; a higher-E BYOK chat can still extract via the private instance.
Add Firecrawl as an Inference Provider (integration Firecrawl), with base URL, optional API key (required for Cloud), and an environment class (E-class). Then add that instance as a webExtract Tool backend.
New workspaces start with an empty backend list — Web Extract stays disabled until you wire Firecrawl.
How it works
- Paste a URL in chat, or ask the model to read a web search result. The model can call Web Extract on that URL.
- Extracted markdown/text is given to the model (truncated at 50,000 characters so one page cannot overflow the turn). You see the URL on the tool card, and sources when present.
- PDFs work when Firecrawl returns text.
- Extracted bytes are not saved as Chat Files.
- If extract fails (timeout, Firecrawl error, blocked URL on their side), the tool card shows an error and the turn continues.
- If no backend is eligible, Web Extract stays in the Tools popover but is disabled with an explanation.
Approval
HITL / HOTL / HOOTL all apply. When approval is required, chat shows the URL the model requested before Firecrawl is called. Denying skips the fetch.
Configure approval under Settings → Native Tools. See Tool policies.
Privacy and providers
The target URL is sent to your Firecrawl instance (Cloud or self-hosted). Steinkauz does not fetch the page origin. Usage and routing follow Data handling. Backend Environment class and provider environment / data routing still apply after backend selection.
Private / on-prem operators can run Firecrawl themselves, or start Steinkauz’s opt-in overlay (not part of the default private stack): copy firecrawl/.env.example to firecrawl/.env, set an API key, then docker compose --env-file firecrawl/.env -f firecrawl/compose.firecrawl.yml up -d (from docker/private/). Add a Firecrawl provider with that base URL (default host port 3100) and API key, then add the Web Extract Tool backend. You operate Firecrawl (including AGPL). Steinkauz is the connector, not the crawler.