Delegate thedeep web work.

Those tend to come back as walls of text you have to dig through. TinyResearch is formatted for decisions: a summary, comparison and decision tables, clear headlines and numbered citations, so you can take it into a meeting without rewriting it. It reads the web through TinyFish’s own search and fetch, writes only from claims it pulled off real pages during the run, and sticks to the sources you set.

A chat assistant can give you a quick table of a dozen rows. TinySet builds a dataset: it splits a big ask into smaller parts, works through each one, and fills every cell from a page it read during the build. You get a dataset you can load into a sheet, a model or a warehouse.

A planner breaks your question into the smaller questions that need answering. Each round, it searches widely, ranks what it finds by credibility and relevance, and reads the best pages live. A separate pass pulls out the specific claims each page supports, with the source attached. The planner checks those findings against your question and, if there are gaps, runs another round. The report is written only from the extracted claims, so every statement traces back to a page.

First it turns your prompt into a plan: what each row is, and the columns to collect for it. You can review the plan and add or change columns before anything runs. Then it splits the job into parts, like dealerships by city, finds the rows in each part, and fills in the remaining fields for every row from real pages. If nothing backs a value, the cell stays empty.

Yes. TinySet shows you the columns it plans to collect before the build starts, and you can change any of them. With TinyResearch you can keep it to sources you trust, block the ones you don’t, limit it to news, research papers or a time window, and review the plan before it starts.

Every value comes from a page TinySet read during the build, and it leaves a cell empty rather than guess. Ask for a source link and a date as columns in your prompt, and anyone can check where a value came from and when.

A dataset is a snapshot of the web at the time it was built. It does not refresh itself, so run it again when you need newer data.

It goes deep, but no tool can promise every result on the web. Coverage depends on how you scope the prompt and what has been published. TinySet builds up to 250 rows per dataset at launch. When something can’t be found, you get an empty cell or a stated gap, never a confident guess.

Yes, so citations alone aren’t the point. The difference is how the report is built: TinyResearch extracts the claims from each page before it writes anything, and writes only from those. That is why there is no filler, and why every line traces back to a source.

Minutes, not seconds. A Deep report typically takes 5 to 20 minutes and a Max report 5 to 45. TinySet takes a couple of minutes to plan and usually 5 to 20 to build. You can run up to 5 reports at once, and each one saves to your dashboard when it’s done.

TinyResearch costs $2 per Deep report and $3 per Max report, paid from your TinyFish wallet. You only pay for completed reports.

In the TinyFish playground, inside ChatGPT and Claude, in agents like Claude Code, Cursor and Codex through MCP, and through the API, CLI and the Python and TypeScript SDKs for teams that build. Both save every run to your TinyFish dashboard.

No. TinySet and TinyResearch read public pages. For data behind a login or paywall, use TinyAgent. For everything from one specific website, TinyFetch or TinyAgent is the better fit.