TinyFish

At a glance

TinyFish is worth evaluating when an agent needs to search the web, read pages and sometimes operate a browser through one provider. Choose the smallest service that answers the question: Search finds pages, Fetch retrieves their content, Browser exposes a controllable session, and Agent attempts a natural-language goal. The main integration work remains checking the answer. A completed agent run does not necessarily mean the requested task succeeded.

Good fit

Developers whose web tasks vary enough that maintaining selectors for every page is costly. A goal can describe the required fields and output shape, while streamed events and saved steps help inspect what the remote agent attempted. For predictable captures controlled by your own code, Browser Run is another managed-browser option. Neither approach removes the need to verify extracted values against the page or respect destination access rules.

Consider something else when

You only need a public page’s text but would pay for an agent to navigate it, or require proof of reliable authenticated automation before evaluation. Search and Fetch have different costs and limits from Agent. Saved login state can expire, sites can demand MFA, and the documented profiles do not solve reCAPTCHA or hCaptcha. The public entry led to signup; no live task, extraction quality or anti-bot performance was established here.

Start here

After choosing to create an account, use the official quickstart’s synthetic shop task: extract the first two product names and prices from scrapeme.live/shop as JSON. The documented raw HTTP route is /v1/automation/run-sse with an X-API-Key header; no SDK installation is necessary for that route. Watch the events, inspect the final result and compare both records with the source page. Handle goal-level failure inside a COMPLETED run as well as infrastructure failure. The example was read, not executed; signup, key creation and paid calls were not performed.

Cost and requirements

On 31 October 2026 wallet pricing listed free Search at 30 requests per minute and 500 per hour, and free Fetch at 150 URLs per minute and 1,000 per day. Agent was $0.016 per step with two concurrent runs by default; Browser was $0.002 per minute with five sessions. New accounts advertised $8 of usage, with a $10 minimum wallet top-up. Search and Fetch work at zero balance, while starting Agent or Browser requires funds. Legacy credit plans use different billing. Step count and retries matter more than the price of one nominal request.

Limits to know

Do not put a password in the goal text: the FAQ says it can appear in logs and model context. Vault connections offer scoped credentials, and Browser Context Profiles save cookies and storage for later runs; these delegate access to the remote service. The claimed separation of credential values from AI planning was not audited. The general privacy policy gives no fixed run-data retention period and allows international processing. Runs can expose steps, screenshots and recordings, so determine retention and access terms before sensitive use. No account, vault connection, saved session or automation was tested.

Related work

• Browser Run: controlled cloud-browser sessions and single-request captures (alternative)

Still unclear

• Task success, extraction accuracy, retry cost and destination compatibility were not measured.
• Contractual retention and access controls for the intended sensitive workload need confirmation before use.