TinyFish, reviewed: one API for web search, extraction, and browser agents

TinyFish gives AI agents one clean way to work with the live web. Search, page fetching, cloud browser sessions, and multi-step agent workflows all sit behind a single API and a single credit pool. If you're a developer or a platform team weighing whether to buy that whole capability instead of wiring it together yourself, this guide walks through what TinyFish does, how the pricing actually works, who's behind it, and the kind of work it's built for.

Key takeaways

• TinyFish bundles four jobs that AI-agent teams normally stitch together on their own: web Search, page Fetch, cloud Browser sessions, and multi-step Agent workflows, all behind one API key and one shared credit pool.
• It's built for agents and applications that need to act on the live web. That means searching fresh results, reading pages that change constantly, logging into sites, and finishing multi-step tasks at production scale.
• The company is serious. It launched in 2025 with a $47M round led by ICONIQ Growth, and the CEO previously ran Nutanix as its President.
• Pricing is credit-based (pay-as-you-go at $0.015/credit, Starter $15/mo, Pro $150/mo, Enterprise custom), where one credit is roughly one agent step or four minutes of browser time. Search and Fetch are free, and failed runs cost nothing.
• It's a strong pick when you want one vendor for the entire "search, extract, act" loop rather than a drawer full of separate tools to maintain.

TinyFish at a glance

What TinyFish is, in one paragraph

TinyFish is a unified web-operations platform built for AI agents and AI applications. The company describes it as infrastructure for "enterprise web agents," software that browses the live web the way a person does to search, extract, and get things done at scale. Its homepage sums it up nicely: "Everything AI needs to use the web." Rather than gluing together a search API, a scraper, a headless-browser host, and your own orchestration code, you call one platform that covers all four. Search returns structured live results, Fetch turns any URL into clean content, Browser gives you cloud browser sessions, and Agent runs multi-step automation. They all share a single API key and one credit pool, and underneath sits Mako, TinyFish's own model tuned for live web work.

The problem it solves will be familiar to anyone who has built an agent. The moment your agent has to touch the real web, the architecture starts sprouting dependencies. You need one provider for search, another to turn HTML into something an LLM can read, a third to drive a real browser through login walls and JavaScript, and then a pile of your own retry, proxy, and anti-bot glue to hold it all together. Every piece has its own SDK, its own bill, and its own way of failing. The web doesn't make it easier, since layouts change weekly and more sites hide behind authentication and bot defenses. TinyFish's bet is that most teams would rather not run that stack themselves. They want an API that goes from lookup to extraction to action in one place, and that keeps working even when a target site has no API of its own.

How TinyFish works: four products, one credit pool

TinyFish routes web work through four coordinated products that share one key and one balance:

• Search delivers structured, live web results (TinyFish markets these as "never cached") that you can feed straight into an agent, instead of scraping a search engine yourself.
• Fetch takes any URL and returns clean, structured, model-ready content in a single call rather than raw HTML.
• Browser provides cloud browser sessions with a stealth browser, anti-bot handling, and an agent proxy, so pages that fight automation still load.
• Agent chains those primitives into multi-step workflows (find, open, extract, act) and returns structured JSON.

Two design choices really matter here. The first is the single credit pool. Every product spends from the same balance, so there's no reconciling four invoices or juggling four rate limits. You top up one meter and use it everywhere, and because Search and Fetch are free, you only spend on the heavier agent and browser work. The second is how TinyFish handles authenticated workflows. Its cloud browser layer is designed to reach the logged-in, gated web without exposing raw credentials to the model, which is exactly the reassurance you want when you're automating anything behind a sign-in.

On reliability, TinyFish is clearly aiming at production rather than demos. It reports 89.9% accuracy on the Mind2Web web-agent benchmark and bills its Agent as "the highest publicly benchmarked web agent," alongside sub-250ms browser cold starts and strong anti-bot pass rates. It also carries SOC 2 and ISO 27001 compliance and offers on-premise deployment for enterprises. These performance figures are TinyFish's own, so the smart move is simply to reproduce them on your own traffic during a trial, which is what any careful buyer would do anyway.

Who's behind TinyFish

For infrastructure you plan to build on, the company matters as much as the API, and TinyFish's story is a reassuring one. It came out of stealth in August 2025 with a $47M round led by ICONIQ Growth, joined by USVP, Mango Capital, MongoDB Ventures, ASG, and Sandberg Bernthal Venture Partners. That's a serious set of backers for a company at this stage. The CEO, Sudheesh Nair, was previously President of Nutanix, and he's joined by co-founders Shuhao Zhang, an engineering leader who came from Meta, and Keith Zhai, a former Wall Street Journal correspondent. The team is based in Palo Alto.

This belongs in a buyer's guide because "who funds and runs it" is a real factor when you're standardizing your web layer on a young platform. Deep funding and an enterprise-operator CEO are good signs that the API will still be here, and improving, a year from now.

Who TinyFish is for

TinyFish fits developers and teams building AI agents that need reliable web search, extraction, and browser automation in one place. Four buyers map neatly onto it:

• Product engineers who want structured, live web data without assembling and babysitting a separate scraping stack.
• Automation teams who need browser-based workflows that can authenticate and finish multi-step tasks, not just fetch a page.
• Data-operations teams who need fresh, typed outputs from constantly changing sites at production volume.
• Enterprise platform teams who need controlled deployment, compliance, and an audit trail for web workflows.

The customer use cases show the shape of the work. For Google Hotels, TinyFish has been used to aggregate inventory from thousands of hotels, surfacing properties that couldn't otherwise be listed, and it did so without asking those hotels to change their own IT systems. For DoorDash, it reportedly collects large volumes of pricing variables to support real-time pricing. TinyFish's own product demos show the same pattern at a smaller scale: an agent that fills quote forms across five separate insurance-carrier portals, clears the bot checks along the way, and returns one structured comparison. Different industries, same job: pull fresh, structured data from a sprawl of external sites that will never build you an API. If your problem rhymes with any of that, TinyFish is aimed squarely at you.

Pricing: do the credit math, not the sticker price

TinyFish is priced in credits across four plans:

Two things make the math friendlier than it first looks. Search and Fetch are free, and failed runs cost $0, so you're billed for successful agent and browser work rather than for retries that never landed. TinyFish defines one credit as roughly one agent step, or four minutes of browser time, with LLM inference, stealth browser, anti-bot, and proxy all bundled in.

That definition is the thing to focus on. The sticker prices are low, so the real question is how many steps your workflow burns per run, multiplied by your run volume. A task that searches, opens several pages, and then acts will naturally use more credits than a one-shot fetch. The easy way to size a plan is to run one representative task on the free credits and count what it costs. You can see the full pricing on the TinyFish pricing page, and the TinyFish listing has the short version. Credits don't roll over on Starter and Pro, so size the plan to your typical volume rather than your busiest day.

TinyFish vs. building it yourself (and vs. point tools)

The most useful comparison isn't TinyFish against one rival. It's TinyFish against the stack most teams assemble by default.

If you want to shortlist by capability, the usual names are Browserbase for managed headless-browser sessions, Firecrawl for turning URLs into clean markdown, Apify for its scraper marketplace, and Bright Data for large-scale proxy infrastructure. Those are mostly raw infrastructure: they hand you a browser, a proxy, or clean text, and you bring the agent logic and the model. TinyFish's angle is different, because it bundles the agent logic and its own model, Mako, in with that infrastructure and puts it all behind one API. When your need is narrow and stable, a point tool is a fine choice. When you want the whole search-extract-act loop from a single vendor, TinyFish is the more natural home.

Strengths

• Real consolidation. Search, fetch, browser, and agent behind one key and one balance removes a genuine class of integration and billing overhead, and free Search, Fetch, and failed runs make it cheap to experiment.
• Built for authenticated, multi-step work. Reaching the logged-in web without handing raw credentials to the model is the detail that separates a "fetch a page" tool from a "finish the task" platform.
• Aimed at production. Anti-bot handling, a stealth browser, an agent proxy, SOC 2 and ISO 27001, and on-prem deployment all point at workloads that have to keep running.
• Serious backing and real customers. A $47M ICONIQ-led round and named users like Google and DoorDash are exactly the signals that make an internal case easier.

What to check before you commit

None of these are dealbreakers. They're the normal diligence you'd do on any web-automation platform:

• Validate the performance numbers on your own traffic. Figures like 89.9% Mind2Web accuracy are TinyFish's own, so run your real sites and tasks on the free credits and confirm they hold up for your use case.
• Model your credit burn. Agent workflows spend a credit per step and credits don't roll over, so meter a representative task before you scale and pick the plan that matches your typical volume.
• Plan for one vendor. Consolidation is the whole appeal, and it also means standardizing your web layer on one platform, so keep your integration reasonably portable, as you would with any core dependency.
• It's a newer platform. TinyFish launched publicly in 2025. The funding and team take a lot of the risk out of that, and it's simply worth noting on a long-horizon bet.

The bottom line

If you're building AI agents that have to operate on the live web, and you'd rather buy the "search, extract, act" loop than build and maintain it, TinyFish is a strong choice. The single API and shared credit pool cut real friction, the authenticated-workflow handling is well thought out, the funding and team are credible, and the enterprise tier answers the questions regulated buyers ask. Do the two pieces of homework that every web-automation purchase deserves: try it on the free credits, and model your credit burn at real volume. For teams that want one vendor for the entire web layer, TinyFish is built for exactly that. You can start on the TinyFish website or read the structured overview on its AgentsIndex listing.

Getting started and further reading

The nice thing is that you can just try it. TinyFish is self-serve, you can create a free account without a card, and 500 free credits is enough to run a handful of real agent tasks and see your credit burn before you pick a plan. A few good starting points:

• The listing. A structured overview of products, pricing, and fit on AgentsIndex.
• Pricing. The current plans and credit definitions live on the TinyFish pricing page.
• Code examples. TinyFish publishes an open-source cookbook of sample apps and recipes on GitHub, which is the fastest way to see what an agent task looks like in code.

The single most useful thing to do before committing is to take one representative workflow, run it end to end on the free credits, and count the credits it consumes. That number, times your expected monthly volume, is your real cost, and it's the figure that tells you whether Starter, Pro, or Enterprise is the right home.

Frequently asked questions

What are the rate limits on each plan?

Search is capped at 60 requests per minute on every plan. Fetch scales with the plan: 150 URLs per minute on pay-as-you-go, 300 on Starter, 600 on Pro, and a custom ceiling on Enterprise. Agent and Browser aren't rate-limited that way at all. They're bounded by concurrency instead: 2, 10 and 50 concurrent agent runs (and 5, 10 and 50 concurrent browser sessions) as you move up the plans.

Is the proxy included in a credit, or billed separately?

Agent runs include proxy automatically, at no extra charge. Browser sessions are the exception: proxy there is metered separately by plan. It's a small distinction that matters if you're budgeting a browser-heavy workload rather than a step-heavy agent one.

Can TinyFish operate sites that require a login?

Yes, and it's a design goal rather than a bolt-on. Browser Context Profiles persist logged-in state between runs, so a recurring workflow doesn't burn steps re-authenticating each time, and credentials live in a vault instead of being pasted into task prompts. TinyFish reports an 85% pass rate against the major anti-bot services it lists (Akamai, Cloudflare, DataDome, PerimeterX and Imperva).

What happens when a run fails, and what does it cost?

Nothing: failed runs are $0 on every plan. That matters more than it sounds, because retries are routine when you're driving third-party sites that change without notice. Browser time is metered at one credit per four minutes, with a 60-minute cap on a single session.

How is Mako different from pointing a frontier model at a browser?

Mako is TinyFish's own web-native model, trained on authenticated, multi-step workflows taken from production runs rather than on general web text, and generally available to enterprises since 21 July 2026. TinyFish reports 89.9% overall on the Mind2Web benchmark, 81.9% on the hard split, and positions Mako at a fraction of frontier-model cost for the same work. Those are vendor-published figures, so treat them as a starting point for your own evaluation rather than a settled result.

What does TinyFish offer teams in regulated environments?

The compliance story sits on the Enterprise plan: SOC 2 and ISO 27001, on-premise deployment, reserved capacity pools for high concurrency, and a dedicated engineer. Pro gets a shared Slack channel and onboarding; pay-as-you-go and Starter are email support.

What are some alternatives to TinyFish?

Depending on the capability you need, common alternatives include Browserbase (managed browser sessions for agents), Firecrawl (clean markdown extraction for LLMs), Apify (a scraper marketplace and run platform), and Bright Data (large-scale proxy and web-data infrastructure). Those are mostly infrastructure you add your own agent logic to, whereas TinyFish bundles the whole search-extract-act loop, and its own model, behind one API.

Sources

• TinyFish, the product homepage and pricing page (product, positioning, and performance claims are as stated by TinyFish)
• TinyFish, the open-source cookbook on GitHub
• ICONIQ Growth, "Powering the future of enterprise web agents", the lead investor's partnership note (funding, mechanism, Google Hotels use case)
• TechStartups, "TinyFish launches from stealth with $47M" (launch, funding, customers, DoorDash use case)
• Reuters via Yahoo Finance, TinyFish $47M raise coverage

Performance and benchmark figures (such as Mind2Web accuracy) and customer references are reported by TinyFish and cited here as vendor statements. Pricing reflects TinyFish's public pricing page at the time of writing and may change.