Agent architecture
The product facts below are checked against the Kosmico source code for each release. They separate what the agent can do, what needs your permission, what differs between desktop and web, and what leaves your device.
A long research request can involve dependent actions, but access to tools, write operations, approvals, iteration limits, and the selected mode constrain the run. A capability shown in one surface should not be assumed to exist identically in every surface.
So this page lists the modes, the permissions and the differences between desktop and web instead of using autonomy as a single marketing label.
Project context can include files, papers, notes, the Research Map, project memory, and recent work. Kosmo reads what the current task needs; the entire project is not pasted into every model call.
Switching models changes the model, not the project that owns the context. A model you pick stays selected; Recommended lets Kosmico pick the model your plan includes before each message.
A project file stored locally, a project with cloud sync on, a cloud-model request, and a local-model request have different data flows. They are listed separately below so local-first is not misread as a promise that every operation is offline.
Official source
Kosmo can chain many steps on its own, but only with the tools of the active mode: a tool outside the mode is refused. Every run has a fixed budget of tool turns, from 10 to 50 depending on where Kosmo runs (20 in the desktop chat, 50 on the web), and ends with a written answer when the budget runs out. Helper sub-agents are read-only and get 5 to 8 turns.
Cloud models come from a catalogue Kosmico manages and are included in your plan, metered as tokens. Free plans get the free models; paid plans unlock the rest. The catalogue is updated from the admin dashboard, so the list can change between releases.
Citation identity
Every citation comes from a record fetched from one of the databases, and you can open it. The review may cite only papers found in the run: a section with an unknown reference, or a substantive paragraph without a citation, is rejected and rewritten. An optional pass also resolves each reference by DOI, arXiv ID, PubMed ID or title and flags the ones it cannot confirm.
Claim-to-evidence validation
A claim taken from a paper's full text is kept only if its quoted passage appears on the cited page; claims without a matching quote are dropped. Each section is written from the evidence of the papers it cites, but no model checks that every sentence of the review is supported, so check the sources of anything you are about to publish.
From the app
Desktop does everything; the web app lacks the parts that need your machine. Desktop only: local files, SSH remote projects, the terminal, Jupyter notebooks, local Ollama models, the MCP server other AI tools connect to, the local research library, and importing from a local Zotero library. Everything else — writing, collaboration, research, the dashboard — works in both.
Settings → AI → Model selection is where you choose the model. It applies to chat, the floating assistant and the chat widget, and you can set how hard the model thinks there too. If you would rather switch mid-conversation, turn on Show in chat on the same screen and a model picker appears in the chat input. Cloud models are served by Kosmico and included in your plan; on desktop there is also a Local section for models you have downloaded through Ollama. Free plans get the free cloud models; paid plans unlock the rest.
No. Kosmico serves cloud models through its own backend and there is nowhere to enter an API key. Usage is metered as tokens against your plan. To spend less, pick a smaller model — or a local one, which is free.
Desktop only. Install Ollama and leave it running at localhost:11434 — Kosmico detects it at startup. Then open Settings → AI → Local AI, pick a model (Kosmico checks your GPU, CPU and RAM and highlights the ones that will run well), and download it. Local models cost nothing, work with no internet, and your data never leaves the machine.
Here is what I can tell you for certain. In a local desktop project your files stay on your own disk; turning on cloud sync or sharing copies their contents to our cloud so collaborators and the web app can reach them. When you chat with me, your message and whatever I read to answer it go to the model you selected — with a local Ollama model, nothing leaves your machine. For anything beyond that — retention, who can access what, whether data is used for training — read the privacy policy at kosmico.ai/privacy or ask support, rather than taking my word for it.
Yours. Terminal commands and notebook kernels run on your own machine, or on the server you connected to over SSH. Kosmico does not execute your code on our infrastructure.
Yes, within limits. In Agent mode it chains steps with the tools the model supports, inside a fixed budget of tool turns. Terminal commands, notebook cells, SSH and Research Map changes ask first unless you give the project Full access, and every file edit appears as a diff you can revert.
No. Kosmo reads what the task needs, and that is what goes to the cloud model you selected. With a local Ollama model on desktop, nothing leaves your machine.
No. Cloud models are served by Kosmico and included in your plan, and there is nowhere to enter an API key. Local Ollama models on desktop cost nothing.