TinyFish Reviews & Product Details

Tinyfish provides exclusive AI agents that can actually surf websites and return structured outputs. They support different trigger modes, including API-triggered and headless modes, as well as geo-specific mode. There is also a silent crawling mode.

These are goal-oriented agents, not just tools. Each search session is available and integratable via streamable APIs, and every session includes logs that can be pulled through separate APIs. If Tinyfish is used within other websites, we can be reassured that we don’t need to host LLMs or build and maintain a system for managing search sessions, history, and related details—Tinyfish takes care of all of that. Parallel sessions are also possible, which helps improve the speed of research. The UI has the sessions, it holds profiles and is sleek. The pricing is optimal. The communities are well present and active for any help and getting started is really easy. Review collected by and hosted on G2.com.

Tinyfish web sessions or instances are independent and do not share information with each other. So, if we need a parent-child relationship between browser sessions and want certain parts of a campaign to run in parallel mode, it becomes difficult. For example, if we want agents to trigger multiple parallel sub-sessions, each with its own browser, while still sharing the data collected across those browsers, that kind of feature is not available. Review collected by and hosted on G2.com.

What I like most about TinyFish is how it lets AI agents interact with real websites and complete tasks through the browser instead of being limited to APIs or static data. This makes it especially useful for automating workflows that would normally require manually navigating websites, extracting information, and taking actions across different web services.

The biggest value for me is the combination of AI intelligence and browser automation. It allows me to turn a natural-language goal into an actionable workflow, which can significantly reduce repetitive work. I also like the potential for integrations because the same approach can be applied across websites without needing to build a custom integration for every individual service.

The overall experience makes TinyFish feel less like a simple scraping tool and more like infrastructure for building practical AI agents that can actually operate on the web. On a daily basis, TinyFish is incredibly smooth to integrate into our automated pipelines. The interface—whether using their REST API, TypeScript SDK, or CLI—is intuitive and cuts out the huge headache of managing browser servers, proxies, and anti-bot bypass mechanisms manually. The sub-250ms cold starts mean our agents execute web actions instantly without hanging. It provides a seamless, developer-friendly experience that removes the clunkiness of old web scraping stacks. Review collected by and hosted on G2.com.

The main area I think TinyFish could improve is the learning curve around building and debugging more complex browser-agent workflows. The basic concept is powerful, but when an agent encounters an unexpected website state or a workflow fails midway, it would be helpful to have even more detailed debugging information and clearer visibility into exactly what the agent attempted.

I’d also like to see broader integrations, more examples for advanced use cases, and clearer guidance around performance and pricing as workflows scale. These improvements would make it easier to move from experimenting with TinyFish to confidently using it in production. Review collected by and hosted on G2.com.

TinyFish is the page-scrape layer I actually trust after a search API. I run Tavily for broad hits, then TinyFish run-sse against a specific URL with a goal — Gensyn docs, DoraHacks briefs, GitHub orgs, hackathon portals — and it returns extractable text instead of “10 blue links.” Header is just X-API-Key. That one-call shape is easy to drop into a Node agent: POST { url, goal }, parse SSE data: lines, keep the COMPLETE payload. On my research desk I credit-gate it (max per cycle/hour) so I don’t burn the key, and I only fire it when a question cites a primary URL. That’s the upside: it closes the loop from “I found a page” to “I have the words on that page,” which is what stops me inventing deadlines and prize amounts. The automated piece I use is the web agent over SSE, not a chat UI. I POST a live URL plus a goal (extract get-started network names, pull prize/deadline text, summarize this GitHub org page) and parse data: events until COMPLETE. That is what sits in my TypeScript desk: credit-gated, max per cycle/hour, only when a question already cites a primary URL. The AI is useful when the goal is tight and the page is official. It is weaker when the goal is fuzzy or the page is a listing dump — you still have to read the extract and mark thin results unverified. I do not let it auto-trade or auto-publish off the scrape. Review collected by and hosted on G2.com.

It isn’t magic on hostile pages. Cloudflare-gated UIs and noisy listing pages (I burned a pass on a Reddit listing that came back generic/unrelated) still fail or waste a credit. I treat a thin extract as unverified, not as a fact. Credits mean you have to budget it; I wouldn’t fire TinyFish on every search hit. Review collected by and hosted on G2.com.

I built my own AI news-monitoring and research tool, and Fetch is the piece that made it viable. I hand it a URL and get back clean Markdown instead of a wall of nav bars, cookie banners and ads, which cuts my token cost on every LLM call downstream. Search and Fetch being free is genuinely rare — everyone else meters exactly this. Integration was trivial: one API key, plain REST, no SDK to adopt. Failed URLs not counting against quota is a small thing that matters a lot when you're pulling from news sites that move or drop content.

Over my last 100 fetches the success rate is 98%, with an average latency of 2.7 seconds. Most pages come back in under two seconds, which matters when I'm processing a batch of feed links in one go.

Onboarding was about as short as it gets: sign up, copy the key, first successful fetch inside five minutes, no credit card in the way. And the output is built for the job — Fetch returns content already shaped for an LLM to read, so the AI layer of my own tool gets a clean input instead of raw HTML it has to fight through.

The dashboard deserves a mention too. Success rate, average latency, extracted content size and credit usage are all on the landing page for each API — I went looking for my failure rate and had it in one click, without digging through logs. For a free tier that's more observability than I expected. Review collected by and hosted on G2.com.

Fetch is unreliable on paywalled and aggregator pages, and unpredictably so. My success rate is 98% over the last 100 fetches, but the failures aren't random: a paywalled article on The Athletic, and a press-review page on an Italian sports aggregator. What makes it awkward to design around is that four other articles from the same publication went through fine in the same window. So I can't route by domain — I have to handle failure per URL and decide at runtime whether a source is usable. Clearer signalling on why a fetch failed would help more than a higher success rate would.

Second, the free tier ends exactly where the platform gets interesting. Search and Fetch cost nothing, Agent and Browser need a funded wallet. There's no middle step, so one stubborn source pushes you from free to pay-per-step.

Related to that, I'm running production tooling on a free tier with nothing stated behind it — no SLA, no published commitment that Search and Fetch stay free. I'd rather see a policy than assume it holds. Review collected by and hosted on G2.com.

Through my brand I serve over 200,000 visitors a month, so I live inside research tools all day. TinyFish has become my default web layer inside Claude Code and OpenCode. The MCP connection was easy to set up, and I was running searches and fetches within minutes. Over the past two to three months it has handled my daily query volume without breaking a sweat. The dashboard is clean, the search results are relevant, and I like that I can jump from a quick search to a full browser session without leaving my coding environment. Review collected by and hosted on G2.com.

The dashboard is clean and easy to use, and the overall experience is solid. The only thing holding it back from a perfect score is the browser automation speed. It works well, but it can be slow and I have hit timeouts on longer verification runs. When it finishes the results are good, but faster and more reliable agent sessions would make a big difference for the amount of daily browser work I do. It is not a dealbreaker because search and fetch still cover most of my needs, but that is the main area I would like to see improved. Review collected by and hosted on G2.com.

I used TinyFish Agent and Fetch in July 2026 to collect a JavaScript-rendered RIDI webtoon bestseller list. Fetch is free and quick. Four pages came back in about 3.2 seconds, which is useful when I just need a first look at a public page. The product split is clear: Search and Fetch do not burn credits, Agent is 1 credit per step, and the docs are honest that COMPLETED means the browser finished, not that the goal succeeded. Review collected by and hosted on G2.com.

On that same RIDI list, Agent looped for about 20 minutes and 127 steps, then failed. A guided page-1 retry returned all 60 titles after about 11 minutes, but it fabricated episode and rating values for ranks 26-48. Episode accuracy was 37/60 and rating/count accuracy was 33/55. One checked row, rank 26, was actually 146 episodes and 28 ratings; Agent returned 100 episodes and 4,500 ratings. Fetch also dropped 28 of 200 titles. I cannot trust Agent output on a virtualized list without independent field checks. Review collected by and hosted on G2.com.

TinyFish is the one layer I actually keep in the stack when I need live web, not another chatbot pretending it can browse. Search, Fetch, Browser, and Agent sit under one API, so I am not wiring Exa for search, Firecrawl for markdown, and Apify for the messy pages. Yes. For how I actually use TinyFish, I’m getting more than I’m paying.

Search and Fetch are free like 30 searches/min, 150 fetches/min, no Wallet draw even at $0. That’s 90% of my work: live recon, clean page content, Reddit/HN/X/LinkedIn listening, lead-gen. I’m not burning credits on the daily path. Review collected by and hosted on G2.com.

Search + Fetch is the sweet spot. Agent is heavier and I do not want it on every listening sweep. MCP is great in Claude/Cursor; connecting it as a custom connector elsewhere (Perplexity-style OAuth) was fiddly. It is also not Exa if I need semantic “find me papers/pages like this,” I still reach for Exa. TinyFish is live web + browser, not neural retrieval Review collected by and hosted on G2.com.

What I like most about TinyFish is how well it handles web-based sources that are hard to reach through conventional APIs. In GraftWatch, it helps enrich ClinicalTrials.gov data by adding India-specific registry information, non-U.S. trial records, and sponsor updates. That extra layer of regional context makes the tool more valuable for researchers who need more than just global trial data. I also appreciate that TinyFish can be used selectively, with budget controls, instead of adding unnecessary cost to every search. Review collected by and hosted on G2.com.

The main thing I dislike is that integrating TinyFish can require more setup and clarification than a conventional API, especially when working with web sources that change their structure. The API contract and response formats also need to be very clear for production use. In GraftWatch, I had to add budget limits and fallback handling so a failed or expensive enrichment request would no Review collected by and hosted on G2.com.

The configuration is very simple and it supports MCP, making it basically ready to use out of the box.

I have integrated web search capabilities into different coding agents and different harnesses. This way when using coding agents, I can have it directly search for the latest and most relevant information, rather than being limited by the training data cutoff date (which often results in outdated content being provided).

Because it is not limited to the services provided by a particular model provider, so it can be used with any harness paired with any open-source model. For me, its biggest advantage is its strong versatility. Review collected by and hosted on G2.com.

Not really, because it's a free service. Review collected by and hosted on G2.com.

This is the easiest way I’ve found to enrich search for my local LLMs, and it does so without any incremental cost or tokens spent. Review collected by and hosted on G2.com.

Integrating it with my local harness was a bit difficult and not as straightforward as I would have liked. As a non-technical founder, I had to dive deep into the docs to get it up and running. Review collected by and hosted on G2.com.