AI brokers wrestle with duties that require interacting with the dwell internet — fetching a competitor’s pricing web page, extracting structured knowledge from a JavaScript-heavy dashboard, or automating a multi-step workflow on an actual web site. The tooling has been fragmented, requiring groups to sew collectively separate suppliers for search, browser automation, and content material retrieval.
TinyFish, a Palo Alto-based startup that beforehand shipped a standalone internet agent, is launching what it describes as the whole infrastructure platform for AI brokers working on the dwell internet. This launch introduces 4 merchandise unified below a single API key and a single credit score system: Net Agent, Net Search, Net Browser, and Net Fetch.
What TinyFish is Delivery
Here’s what every product does:
- Net Agent — Executes autonomous multi-step workflows on actual web sites. The agent navigates websites, fills types, clicks by flows, and returns structured outcomes with out requiring manually scripted steps.
- Net Search — Returns structured search outcomes as clear JSON utilizing a customized Chromium engine, with a P50 latency of roughly 488ms. Opponents on this area common over 2,800ms for a similar operation.
- Net Browser — Offers managed stealth Chrome periods through the Chrome DevTools Protocol (CDP), with a sub-250ms chilly begin. Opponents usually take 5–10 seconds. The browser consists of 28 anti-bot mechanisms constructed on the C++ degree — not through JavaScript injection, which is the extra widespread and extra detectable method.
- Net Fetch — Converts any URL into clear Markdown, HTML, or JSON with full browser rendering. Not like the native fetch instruments constructed into many AI coding brokers, TinyFish Fetch strips irrelevant markup — CSS, scripts, navigation, adverts, footers — and returns solely the content material the agent wants.
The Token Downside in Agent Pipelines
One of many constant efficiency issues in agent pipelines is context window air pollution. When an AI agent makes use of a regular internet fetch instrument, it usually pulls the whole web page — together with hundreds of tokens of navigation components, advert code, and boilerplate markup — and places all of it into the mannequin’s context window earlier than reaching the precise content material.
TinyFish Fetch addresses this by rendering the web page in a full browser and returning solely the clear textual content content material as Markdown or JSON. The corporate’s benchmarks present CLI-based operations utilizing roughly 100 tokens per operation versus roughly 1,500 tokens when routing the identical workflow over MCP — an 87% discount per operation.
Past token rely, there’s an architectural distinction value understanding: MCP operations return output immediately into the agent’s context window. The TinyFish CLI writes output to the filesystem, and the agent reads solely what it wants. This retains the context window clear throughout multi-step duties and allows composability by native Unix pipes and redirects — one thing that’s not potential with sequential MCP round-trips.
On complicated multi-step duties, TinyFish stories 2× larger process completion charges utilizing CLI + Expertise in comparison with MCP-based execution.
The CLI and Agent Talent System
TinyFish is delivery two developer-facing elements alongside the API endpoints.
The CLI installs with a single command:
npm set up -g @tiny-fish/cli
This offers terminal entry to all 4 endpoints — Search, Fetch, Browser, and Agent — immediately from the command line.
The Agent Talent is a markdown instruction file (SKILL.md) that teaches AI coding brokers — together with Claude Code, Cursor, Codex, OpenClaw, and OpenCode — how one can use the CLI. Set up it with:
npx abilities add https://github.com/tinyfish-io/abilities –skill tinyfish
As soon as put in, the agent learns when and how one can name every TinyFish endpoint with out guide SDK integration or configuration. A developer can ask their coding agent to “get competitor pricing from these 5 websites,” and the agent autonomously acknowledges the TinyFish talent, calls the suitable CLI instructions, and writes structured output to the filesystem — with out the developer writing integration code.
The corporate additionally notes that MCP stays supported. The positioning is that MCP is fitted to discovery, whereas CLI + Expertise is the really helpful path for heavy-duty, multi-step internet execution.
Why a Unified Stack?
TinyFish constructed Search, Fetch, Browser, and Agent fully in-house. It is a significant distinction from some rivals. For instance, Browserbase makes use of Exa to energy its Search endpoint, that means that layer shouldn’t be proprietary. Firecrawl gives search, crawl, and an agent endpoint, however the agent endpoint has reliability points on many duties.
The infrastructure argument shouldn’t be solely about avoiding vendor dependencies. When each layer of the stack is owned by the identical workforce, the system can optimize for a single consequence: whether or not the duty accomplished. When TinyFish’s agent succeeds or fails utilizing its personal search and fetch, the corporate will get end-to-end sign at each step — what was searched, what was fetched, and precisely the place failures occurred. Corporations whose search or fetch layer runs on a third-party API wouldn’t have entry to this sign.
There’s additionally a sensible value that groups integrating a number of suppliers encounter. Search finds a web page the fetch layer can not render. Fetch returns content material the agent can not parse. Browser periods drop context between steps. The result’s customized glue code, retry logic, fallback handlers, and validation layers — engineering work that provides up. A unified stack removes the element boundaries the place these failures happen.
The platform additionally maintains session consistency throughout steps: identical IP, identical fingerprint, identical cookies all through a workflow. Separate instruments working independently seem to a goal web site as a number of unrelated shoppers, which will increase the chance of detection and session failure.
Key Metrics
Key Takeaways
- TinyFish strikes from a single internet agent to a four-product platform — Net Agent, Net Search, Net Browser, and Net Fetch — all accessible below one API key and one credit score system, eliminating the necessity to handle a number of suppliers.
- The CLI + Agent Talent mixture lets AI coding brokers use the dwell internet autonomously — set up as soon as and brokers like Claude Code, Cursor, and Codex routinely know when and how one can name every TinyFish endpoint, with no guide integration code.
- CLI-based operations produce 87% fewer tokens per process than MCP, and write output on to the filesystem as an alternative of dumping it into the agent’s context window — conserving context clear throughout multi-step workflows.
- Each layer of the stack — Search, Fetch, Browser, and Agent — is constructed in-house, giving end-to-end indicators when a process succeeds or fails, a knowledge suggestions loop that can not be replicated by assembling third-party APIs.
- TinyFish maintains a single session id throughout a complete workflow — identical IP, fingerprint, and cookies — whereas separate instruments seem to focus on websites as a number of unrelated shoppers, rising detection threat and failure charges.
Getting Began
TinyFish gives 500 free steps with no bank card required at tinyfish.ai. The open-source cookbook and Talent information can be found at github.com/tinyfish-io/tinyfish-cookbook, and CLI documentation is at docs.tinyfish.ai/cli.
Observe: Thanks for the management at Tinyfish for supporting and offering particulars for this text.
