What is GPT Researcher? Is it free, how to install it, Docker, Ollama, MCP and license

You give GPT Researcher a topic; it researches the web (and, if you want, your own documents) and writes a report with a source list. It is an open-source AI agent; you install it on your own computer with Docker or Python, with no account to create. It needs your own language model key (OpenAI by default) and search key (Tavily by default), and you pay those providers. Through the MCP server in a separate repository, it connects to Claude Desktop as a research tool. The official sources don't promise Turkish support.

What it does: Researches a topic on the web and writes a cited report
Free: Software yes · API keys are on you
No-code?: No, it needs setup
Setup: Docker or Python 3.12+
License: Apache 2.0 / MIT (conflicting)
Turkish: Not stated in the official sources
Version: v3.7.0 · September 26, 2026

What does GPT Researcher do? Can I use it without code?

Can I use it without code?

There is no fully no-code path. There is no official cloud product or account; you install it yourself. With Docker, you type a few commands in the terminal and fill in a settings file (.env); after that, you use the interface in your browser.

You write a research question. GPT Researcher first splits it into several sub-questions, reads sources for each one in parallel on the web (or in your own documents), filters out the useful parts and writes a report with citations. So instead of asking a chatbot one question at a time, you get a long piece of preliminary research done in one go.

PlannerSplits the question into research questions.

Reader agentsRead sources in parallel and summarize them.

ReportFilters the findings and turns them into a cited report.

The report cites sources, but the README describes the project as "experimental, as is": check the claims against the sources.

Get a cited research reportWhat you get: Cited reportNeeds a separate setup Install with DockerWhat you get: A working interfaceNeeds a separate setup Add a research tool to Claude (MCP)What you get: deep_research, quick_search, write_report toolsNeeds a separate setup

Who it's for: consultants, content creators, students and developers who want to do market, competitor or topic research on their own computer with a language model of their choice. Someone who doesn't want to deal with code may find a ready-made deep research app easier.

Source: GitHub README and the gptr.dev home page; October 15, 2026.

How do you install GPT Researcher? Docker, Python and pip

There are three ways: Docker, running the repository with Python, or the pip package. The first two give you a browser interface; the pip package is for people who write code. All of them need two API keys: one from your language model provider (OpenAI by default) and one from your search provider (Tavily by default). An API key is a secret password you get from the provider's dashboard; don't share it with anyone and don't write it into code files.

Install and open Docker Desktop.
Download the repository: use the commands below and enter the folder.
Prepare the settings file: copy .env.example, save it as .env, and fill in OPENAI_API_KEY and TAVILY_API_KEY.
Start: docker compose up --build. If that doesn't work, try it with a hyphen: docker-compose up --build.
Open: The server starts at localhost:8000 and the interface at localhost:3000.

1. Download the repository and enter the folder

› git clone https://github.com/assafelovic/gpt-researcher.git
cd gpt-researcher

2. Start› docker compose up --build

If you will use a language model other than OpenAI, the docs say to add that provider's langchain package to requirements.txt.

Install Python 3.12 or later.
Download the repository (the git clone command above) and enter the folder.
Provide the keys with export or with a .env file.
Install the dependencies, start the server and open localhost:8000.

Keys (set on Windows, or a .env file)

› export OPENAI_API_KEY=your-key
export TAVILY_API_KEY=your-key

Install and start

› pip install -r requirements.txt
python -m uvicorn main:app --reload

For use inside your own Python code. The README example: the GPTResearcher class takes the query, runs the research and writes the report.

Install› pip install gpt-researcher

README example (summary)

› from gpt_researcher import GPTResearcher
researcher = GPTResearcher(query="...")
await researcher.conduct_research()
report = await researcher.write_report()

The Python version differs across sources: the README and gptr.dev say 3.11+, but pyproject, PyPI, the installation docs and the v3.7.0 release notes require 3.12+. Install 3.12 or later. If you install with Docker, the Python version on your computer doesn't matter.

Account and hardware

There is no GPT Researcher account. Hardware requirements (RAM, disk, GPU) are not stated in the official sources; if you use a cloud language model, the heavy lifting happens at the provider. If you use a local model (Ollama), that model's own requirements apply.

Current version

v3.7.0 on GitHub (September 26, 2026); 0.16.1 on PyPI (September 26, 2026). To update a Docker or repository install, update the repository (git pull); for pip, run pip install -U gpt-researcher.

Source: repository docs (getting-started, Docker), README, pyproject.toml, PyPI; October 15, 2026.

How do you use GPT Researcher? Your own documents, Ollama and search engines

Open the interface. localhost:8000 (localhost:3000 with Docker).
Write your research question. Say what you want to learn and from which angle.
Choose the source type. Under Report Source, pick web or My Documents (your own documents).
Read the report. The report comes with a source list; export to PDF and Word is mentioned in the README.

Sample research question"Use of AI-powered customer service on e-commerce sites: what developments have there been in the past year, and which problems stand out?"The report language is English by default and is set with the LANGUAGE setting. The quality of Turkish reports is not promised in the official sources; you need to try it.

Researching your own documents (PDF, Word, Excel, PowerPoint...)

The README lists the supported formats: PDF, plain text, CSV, Excel, Markdown, PowerPoint and Word. You point to the folder holding your documents with the DOC_PATH variable; in the interface, you choose My Documents under Report Source.

Point to the documents folder› export DOC_PATH="./my-docs"

Document text still goes to the language model provider you chose. For confidential documents, the data leaves your computer unless you connect a local model.

Changing the search engine (Tavily, DuckDuckGo, SearXNG, arXiv...)

The default search provider is Tavily. You change it with the RETRIEVER variable, or combine several separated by commas. The docs list these options: Tavily, Bing, Brave, Google, Serper, Searx (SearXNG), DuckDuckGo, arXiv, OpenAlex, Semantic Scholar, Exa, PubMed Central and others. Each provider has its own key requirement.

Combine web and academic sources› RETRIEVER=tavily,arxiv

Running with a local model (Ollama)

The docs give an example of connecting a local language model with Ollama: in the .env file, you set FAST_LLM, SMART_LLM and STRATEGIC_LLM to a model name starting with ollama:, and EMBEDDING_PROVIDER to ollama. The example in the docs uses a small test model; there is no official promise about report quality.

Format of the example settings in the docs (choose the model name yourself)

› OLLAMA_BASE_URL="http://127.0.0.1:11434/"
FAST_LLM="ollama:model-name"
SMART_LLM="ollama:model-name"
STRATEGIC_LLM="ollama:model-name"
EMBEDDING_PROVIDER="ollama"

LM Studio: the official docs don't mention LM Studio. The README only describes support for a custom OpenAI-compatible address (OPENAI_BASE_URL); whether it works with LM Studio is not clear in the official sources.

Deep research mode

It branches the topic into subtopics like a tree and researches each branch separately. The docs say a deep research run takes about 5 minutes and cost about $0.40 with an older model setup. Because the current default model is different, today's cost isn't in the docs.

Deep research in Python› researcher = GPTResearcher(query="...", report_type="deep")

Source: README, repository docs (search-engines, running-with-ollama, deep_research), default.py; October 15, 2026.

How do you connect GPT Researcher to Claude via MCP? How the skill differs

MCP is the standard for adding outside tools to AI apps. GPT Researcher's MCP server lives in a separate repository (gptr-mcp): once installed in a client such as Claude Desktop, the assistant can run deep research for you. The gptr.dev home page gives Claude, Cursor and n8n as examples.

GPT ResearcherClaude DesktopInstalled from the gptr-mcp repositoryCursor, n8n, other MCP clientsMentioned on the home page; no detailsYour own documentsVia DOC_PATHOther MCP serversRETRIEVER=tavily,mcp

Tool	What it does
deep_research	Researches a topic in depth
quick_search	Speed-focused web search, with summaries
write_report	Writes a report from research results
get_research_sources	Returns the sources used in the research
get_research_context	Returns the full research context

Claude Desktop setting (claude_desktop_config.json, format from the README)

› { "mcpServers": { "gptr-mcp": { "command": "python", "args": ["/absolute/path/gptr-mcp/server.py"], "env": { "OPENAI_API_KEY": "your-key", "TAVILY_API_KEY": "your-key" } } } }

According to the gptr-mcp README, the server is slower than a normal search (about a 30-second wait). The repository is MIT-licensed; its last update was November 7, 2025, which is older than GPT Researcher's latest releases. The official sources don't separately state compatibility with the new version.

The Claude skill doesn't run research

The README says it "extends Claude's deep research capability," but the skill file in the repository is aimed at people developing GPT Researcher: understanding the architecture, adding features, debugging. Install: npx skills add assafelovic/gpt-researcher. If you want Claude to run research, install the MCP server above.

Source: gptr-mcp README, SKILL.md in the repository, claude-skill docs, gptr.dev; October 15, 2026.

Is GPT Researcher free? What does it really cost?

The GPT Researcher software is open source and free: no subscription, plan or account. But it needs keys for outside services, and you pay their providers yourself. The software is free; the APIs you use, a server (if any) and the work of keeping it updated are on you.

Item	Who charges	Cost
GPT Researcher software	Open source	Free
Search (Tavily by default)	Tavily	Tavily's own docs: 1,000 free credits a month (no card required); paid monthly plans and pay-as-you-go are available. How many credits one research run uses is not in the GPT Researcher docs
Language model (OpenAI by default)	Your provider	Paid by usage; the amount is not in the sources, see the provider's pricing page
Local model (Ollama)	None	No API fee; the hardware to run it is yours
Server (optional)	Your hosting company	If you install it on your own server, that server's cost; no amount given

The official cost figures are old

The introduction docs put an average research run at about 3 minutes and $0.10, and a deep research run at about 5 minutes and $0.40. Both were measured with old model settings; the current default models in the configuration file are different. You need to work out today's real cost from your providers' pricing pages.

Cutting costs

MAX_ITERATIONS (default 3) sets the number of search rounds
MAX_SEARCH_RESULTS_PER_QUERY (default 5) sets how many results are read per question
TOTAL_WORDS (default 1200) sets the report length
In deep research, the breadth and depth settings increase the cost

Source: README, introduction docs, deep_research docs, default.py and .env.example, Tavily docs (docs.tavily.com/documentation/api-credits); October 15, 2026.

GPT Researcher license, data, security and Turkish

License: two sources say two things

The LICENSE file in the repository, GitHub and the README say Apache 2.0. pyproject.toml and the PyPI page say MIT. Both are permissive licenses, but the repository doesn't say which one applies. Before using it in a commercial project, check with the project owner or the LICENSE file; this is not legal advice. The README also describes the project as experimental and without warranty.

Where does the data go?

The software runs locally, but your research questions, the text of the pages it reads and (if you use them) your own documents are sent to the language model and search provider you chose; those providers' privacy rules apply. Only if you connect a local model does the language model data stay on your computer. If you enter the optional TYPESAFE_API_KEY, the fetched chunks are scored by the TypeSafe service for context filtering; if you don't, local keyword ranking is used.

Where are reports stored?

The Docker setup mounts the outputs, logs and my-docs folders to your computer, so reports stay on your own disk.

Security: don't expose the server to the internet

SECURITY.md says the backend has no authentication by design and should run locally or behind network security you control. The MCP configuration can execute commands; don't expose this endpoint to untrusted people. No security advisories appear on GitHub (October 15, 2026).

Turkish reportsUnclear	There is a LANGUAGE setting that picks the report language (default english). Valid values and Turkish quality are not stated in the official sources.
Turkish interfaceUnclear	The official sources say nothing about the interface language. The README is translated only into Chinese, Japanese and Korean.
Local model (Ollama)Yes	The docs include a setup example; data stays on your computer.
Research from your own documentsYes	PDF, Word, Excel, PowerPoint, CSV, Markdown and plain text.

Source: LICENSE, README, pyproject.toml, PyPI, SECURITY.md, docker-compose.yml, v3.7.0 release notes; October 15, 2026. Model training and data deletion rules depend on the provider, so they are not covered here.

GPT Researcher common problems and fixes

What happens	What to do
"cannot load library gobject-2.0-0" or "pango" error in PDF output	PDF output uses WeasyPrint. On macOS, run brew install glib pango; on Linux, install the libglib2.0-dev and libpango-1.0-0 packages. If that isn't enough on an Apple M-series chip, the docs suggest installing Python 3.12 with brew and running the app with it.
It won't read an intranet, staging or local address	Private and local addresses are blocked by default (SSRF protection). If you intentionally research internal addresses, add ALLOW_PRIVATE_URLS=true to the .env file.
I use FireCrawl, and some sources silently drop out of the report	The FireCrawl free plan allows 2 concurrent browsers; anything beyond that comes back empty without an error. Lower FIRECRAWL_CONCURRENCY and MAX_SCRAPER_WORKERS to fit your plan.
"Error processing the url" or a Chrome version mismatch	According to the docs, some sites can't be read; restart and try again. The docs also describe downgrading Chrome, but that section may be out of date.

Source: the repository's troubleshooting.md and .env.example; October 15, 2026. All entries come from the official documentation.

Frequently asked questions

How is GPT Researcher different from Perplexity and ChatGPT deep research?

The official sources don't publish a comparison, so no claim of superiority is made here. The structural difference: GPT Researcher is not an app but open-source software you install yourself. You choose the language model and the search provider, and you pay the providers. For the other products' features, see their own pages.

How do you install GPT Researcher, and does it run with Docker?

Yes. You download the repository, put your OpenAI and Tavily keys in the .env file, then run docker compose up --build; the interface opens at localhost:3000. Installing with Python requires 3.12 or later.

Does GPT Researcher connect to Claude via MCP?

Yes, the gptr-mcp server in a separate repository adds tools such as deep_research, quick_search and write_report to clients like Claude Desktop. It needs the same API keys. The Claude skill, on the other hand, is for developers and doesn't run research.

Does GPT Researcher work with a local model through Ollama or LM Studio?

The official docs include a setup example for Ollama. LM Studio is not mentioned in the official docs; the README only describes support for a custom OpenAI-compatible address, and whether it works with LM Studio is not clear. There is no official promise about report quality with a local model.

Is GPT Researcher free?

The software is free. You pay the providers of the language model (OpenAI by default) and search (Tavily by default) keys; Tavily's own docs list 1,000 free credits a month. With a local model such as Ollama, you can bring the language model cost down to zero.

What is GPT Researcher's license, and can it be used commercially?

The LICENSE file in the repository is Apache 2.0, but pyproject.toml and PyPI say MIT. Both are permissive licenses; the repository doesn't say which one is definitive. Before commercial use, check with the project owner or the LICENSE file. This is not legal advice.

Tasks and setup kits that use GPT Researcher

Researching competitors and the market

Related guides

Sources

Not clear in the official sources: Hardware requirements · Turkish report quality and LANGUAGE values · interface language · the real cost of one research run with the current default model · working with LM Studio · which license applies (Apache 2.0 or MIT) · gptr-mcp compatibility with the latest version · whether the 'hosted beta' line in troubleshooting.md is current. Each is noted in the relevant section along with what you can do.