How is Research Agent different from Xedant Agent?
Xedant Agent is a developer tool: an IDE around Claude Code for writing code. Research Agent is an analyst’s workspace: market research, planning, and documentation with no coding involved. Research Agent can also control the agent in the background — agent chats and task launches are visible in the same interface.
How is it better than Perplexity or ChatGPT with search?
Those answer a question from a few sources in a single request. Research Agent runs dozens of searches, downloads and reads hundreds of pages, and extracts data in batches — volumes that do not fit into a model’s context. The result is a verified report with primary sources, the kind you cannot assemble by hand.
How much does a single research run cost?
With the DeepSeek-V4-Flash model — roughly $0.09–0.27, finished in 20–30 minutes; DeepSeek-V4-Flash-Mini through OpenRouter costs even less. Any OpenAI-compatible model works: pick any other one, and token and cost statistics are tracked for every call.
What do I need to run it?
A Docker host — one container, the SQLite database is built in. After launch, set the API key of at least one search provider and one LLM service in the Services section. For Xedant Agent integration — the agent’s URL and API key in environment variables.
Can it be used without Xedant Agent?
Yes. Search, downloading, extraction, the research library, and reports all work on their own. The integration only adds agent chats and the “Call Agent” buttons on document pages.
Where is the data stored?
Everything is self-hosted in the /data volume: the database (SQLite or PostgreSQL), downloaded Markdown, keys. The only outgoing traffic is requests to the search and LLM providers you choose.
Why not open-source?
We tried running and supporting free products more than 20 years ago. It is not worth it — the support load is much higher, and no resources are left for further development. We prefer far fewer users who genuinely love the products and are ready to pay for them. One more important reason is control: we like keeping the products simple and minimalistic. With open source, we would expect tons of AI-generated pull requests with features we deliberately avoided because of their complexity and poor design decisions.
What is the prompt queue and why do you need it?
A prompt is a text task for the agent: “collect competitor data” or “find and describe the strengths of five models”. Normally each task has to be written out and sent by hand. The prompt queue lets you sketch out a list of tasks in advance, and the agent runs them itself — one after another, with no manual sending. Turn on AutoSend and several tasks run at the same time: as soon as one finishes, the next one starts. The main point: you prepare the list once, then the agent works on its own — you can get on with other things. The queue is shared with the Xedant Agent interface: tasks added there are visible here too.
Is there voice input?
Yes. Next to the message input there is a microphone button: a long task can simply be dictated — speech recognition works right in the browser, and the dictated text appears in the message and goes to the agent. No typing needed, and you can even dictate tasks into the prompt queue.
Is it safe to keep keys in environment variables?
Yes. In a service’s settings, the secret can be stored not as the key itself but as a reference to an environment variable — $NAME or ${NAME}. The key itself is not kept in the database: the value is read from the server’s environment on every use. To change the key, change the variable’s value and restart the container — no database editing needed.
Can it search Yandex?
Yes. Among Research Agent’s search providers is JsonSeo — it can return results from Yandex, Google, or Bing. Which search engine to use is chosen in the service’s settings, and a search region can be set there as well. This matters when you research markets that Yandex covers better than other search engines. Details in the Search section.
What if a site cannot be downloaded?
Downloading follows a chain of methods: a plain request, a request through a proxy, a local browser, a local undetectable browser, and a remote browser. If a page will not come through one way, Research Agent tries the next one itself — most of the time there is nothing to do. If nothing helped, the reason is usually that the needed method is not configured: check the proxy and the remote browser address in the Services section, or enable Nodriver — a local undetectable browser that downloads everything it needs on first use. Details in the Fetching Pages section.