Web agents that finish the job and cost a fraction per step
CrawlGenie web agents are goal-driven: a web agent takes a goal ("find the cheapest Team plan and its seat limit"), a URL and an optional result schema, drives a real browser, and returns structured JSON. Every step reads a ranked snapshot of a few thousand tokens instead of a page dump, so the same task costs a fraction of what page-pasting agents spend. Run one, or run five hundred in a batch.
How do I run a web agent?
For developers: the exact call
POST /v1/agent/run
{"task": "Find the Team plan price per seat and the seat limit", "url": "https://acme.com/pricing",
"schema": {"plan": "string", "price_per_seat_usd": "number", "seat_limit": "number", "source_url": "string"}}
→ 202 { id, statusUrl } then GET /v1/agent/runs/{id}
→ { status: "done", result: { plan: "Team", price_per_seat_usd: 18, seat_limit: 200, source_url: "…" }, steps: 6, usage: { inputTokens: 9120, observationTokens: 7400 } }How do I run hundreds at once?
For developers: the exact call
POST /v1/agent/batch
{"task": "Return the first available delivery date for postcode 10115", "urls": ["https://shop-a.com", "https://shop-b.com", "…"], "schema": {"date": "string", "fee_eur": "number"}, "concurrency": 5}
→ { batchId, count: 180, statusUrl } GET /v1/agent/batches/{batchId} → done/running/failed counts, per-URL results, total usageBatches respect your concurrency cap and the workspace's browser limit; each item is a normal run with its own timeline, so a failed URL can be replayed on its own.
Why are the steps cheaper than other agent APIs?
Three reasons. The agent reads an interactive, ranked accessibility snapshot (3–8k tokens) rather than HTML or screenshots on every step. When it needs detail it asks for the page's main text once, focused, instead of scrolling screenshots. And anything that does not need a model (lists, prices, jobs, filings) is routed to the 61 built-in agents, which cost zero model tokens. The receipt on every run shows input, output and observation tokens.
What about logins, payments and bot walls?
Logins come from profiles (saved sessions) or the secrets vault: the agent types {{secret:name}} and the server substitutes it, so the model never sees a password. Anything that looks like pay, send or delete waits for a human approval if you turn the gate on. Challenges stop the loop instead of burning steps; the signed agent identity lets supporting sites recognise you. Residential proxies are bring-your-own.
Which model does the agent use?
Yours. Store an OpenAI, Anthropic, Gemini or OpenAI-compatible key once as the workspace secret llm_api_key (or pass it per request). CrawlGenie does not resell model tokens and does not mark them up; you pay only the flat plan.
How does this compare with TinyFish, Browser Use or Operator-style agents?
TinyFish sells agent steps at a per-step price on its own models and infrastructure; Browser Use is a framework you host and feed with page dumps; Operator-style agents run on screenshots. CrawlGenie keeps the agent, the reading layer, the extractors, documents and memory behind one key, lets you choose the model, and meters what each step read. Where a task is pure extraction, it should not need an agent at all, and CrawlGenie will tell you which built-in agent to use instead.
Related: web scraping for LLMs · 61 built-in agents · security model · API reference.
What would this save you?
Move the sliders. The saving is computed from the measured ratios above and the model price you pick.
Assumes 10% of the page is kept per read (measured range 5–15%). Receipts in the product show the real number for your traffic.
Questions people ask
Does the agent return JSON in a fixed shape?
Yes. Pass a schema (a JSON Schema or an example object); the agent is instructed to return exactly that shape in its result, and the run records it under result.
How many agents can run at once?
As many as your plan's concurrent browsers allow; a batch takes a concurrency cap (1–10) and queues the rest. Each item is its own run with a timeline and replay.
Can the agent use my logged-in accounts?
Yes, through profiles (saved cookies and storage) and vault secrets. Approvals can gate purchases, messages and deletions.
What does a step cost?
The plan is flat; the only variable cost is your model's tokens, and the receipt shows them per step. Observations are 3–8k tokens instead of 50k+ for page dumps.
Spend tokens on answers, not on pages.
Create a workspace in a minute. Connect ChatGPT, Claude, Cursor or your own agents. Every answer comes with a token receipt.