Do I need to host GPT Researcher myself?

Yes. It is a connected tool: you run the official GPT Researcher FastAPI backend and point your bot at it over HTTPS. The backend does its own model and search calls, so it carries its own provider configuration and costs.

Why do reports take so long?

Because it is actually reading sources rather than answering from memory. Start reports in the background and collect them from the reports list instead of holding a request open.

How is this different from web search?

Web search answers a question in one hop. GPT Researcher plans a question, gathers and synthesises multiple sources, and returns a cited report you can audit and hand to someone else.

← Home

Guide

GPT Researcher

Ask a language model a research question and you get an answer shaped like research: fluent, plausible, and impossible to check. GPT Researcher does the other thing — it plans the question, goes and reads sources, and gives you a report with citations you can follow. Connects to OpenClaw and Hermes alike.

What It Actually Does

It is a deep-research agent, Apache-2.0 licensed. Given a task it breaks the question down, gathers evidence from the web or from local documents, synthesises what it found, and produces a written report that cites its sources. The output is meant to be audited, which is the entire difference from a model answering from memory.

You Host the Backend

This is a connected tool, not a local install: you run the official GPT Researcher FastAPI backend yourself and point your bot at it over HTTPS. Two things follow that are easy to miss.

The backend does its own model and search calls, so it carries its own provider configuration and its own running costs — separate from the model your bot is on.
A real report takes minutes, not seconds, because it is genuinely reading things. Plan for that rather than treating a slow reply as a fault.

Connect It

Stand up the GPT Researcher FastAPI backend and put it behind HTTPS.
Open the dashboard, go to Tools, and pick a running bot.
Find GPT Researcher and connect it, entering your backend’s address and whatever authentication you put in front of it.

It connects as a plain HTTP tool. Starting a report is a post /report/ with the task and the report options; the shape the catalog ships with sets report_type, report_source and a tone, and can set generate_in_background so the call returns immediately. Finished work is listed by get /api/reports.

That background-plus-list pattern is the one to prefer. A research run is long enough that holding a request open for it is the wrong shape: start it, let the bot come back for the result, and nothing times out in between.

What to Ask For

“Research how our three main competitors price, and cite each claim.”
“What has actually been published on this technique in the last year?”
“Build me a briefing on this company before tomorrow’s call.”
“Start that research in the background and tell me when the report is ready.”

Give it a question with a checkable answer. Its advantage is evidence, so a task that wants an opinion wastes it.

Where It Sits Beside Your Other Tools

Your bot already has web search, and tools like Jina MCP read individual pages cleanly. Those answer a question in one hop. GPT Researcher is for the case where one hop is not enough and you need the working shown — a report you could hand to someone else and defend.

Troubleshooting

Requests time out. Expected on a synchronous report. Generate in the background and fetch the result afterwards.
Reports come back thin. Usually the backend’s own search or model configuration, not the connector. Test the backend directly with the same task.
The bot answers from memory instead. Ask explicitly for a researched, cited report — otherwise a capable model will simply reply.

Try It

Connect it and commission one report on a subject you already know well. Reading a report whose answer you can check is the fastest way to learn how much to trust the ones you cannot.