Workflow comparison

One project context or a chain of handoffs

A conventional research stack can be powerful because each specialist tool is mature. Its cost is the manual work required to move sources, context, decisions, files, and citations between them.

Kosmico's design bet is that research improves when specialist surfaces share one project state and the agent can act across them with explicit permissions.

Specialist tools remain valid choices

Zotero, Overleaf, notebooks, academic databases, and team chat each solve important problems. A team may prefer them for established workflows, institutional integrations, or a specific specialist capability.

The integration cost should be measured

Manual actions, application switches, duplicated context, stale copies, recovery after interruption, and artifact completeness are benchmark outcomes, not cosmetic details. The controlled benchmark holds the model and corpus constant to measure that workflow layer.

Criterion	Kosmico	Conventional stack
Context	One project context selected for the task	Context copied or reconnected across applications
Artifacts	Literature, notes, code, references, and manuscript in one project	Artifacts remain in their specialist tools
Automation	Agent actions across supported project surfaces and permissions	Integrations, exports, scripts, and manual handoffs
Control	One workspace policy and project permissions	Independent permissions, storage, and policies per service

Official source

Current product facts

Kosmico v0.7.6 · Verified October 20, 2026

Autonomy boundary

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.

Agent modes

Agent: The default. Kosmo carries out the task with the tools the model supports: it reads and edits files and, on desktop, runs commands and notebook cells.
Plan: Turned on with /plan. Kosmo reads the project and proposes the steps; nothing in the project changes until you accept the plan, which then runs in Agent mode.
Chat: Used when the selected model has no tool support. Kosmo discusses your work and changes nothing.

Permissions and safeguards

Terminal commands, notebook cells, SSH and Research Map changes ask first: Allow once, Always allow or Deny. Always allow covers one kind of command (for example git) for the rest of that chat.
Commands such as rm -r, sudo or dd never offer Always allow.
File edits do not interrupt the run: each one appears as a diff you can accept, reject or revert.
A project can be switched to Full access, which skips the prompts, including those for risky commands. It applies to the open project only and resets when you leave or restart.
Commands and notebook kernels run on your machine or on the server you connect over SSH, never on Kosmico's servers.
External AI tools connected through Kosmico's MCP server need your approval for each request to your personal memory.

Model policy

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.

Recommended: Recommended is the default: before each message Kosmico picks the model your plan includes. A model you choose yourself stays selected.
BYOK: Not supported. There is nowhere to enter an API key, and a ChatGPT or Claude subscription cannot power the chat. It works the other way round: Claude Code, Cursor and other MCP clients can use your project through Kosmico's MCP server on desktop.

Models by plan

Data flows

Local project (desktop)

Files stay on your disk.
A project you never sync does not leave the machine and does not count toward your plan's project limit.

Synced project

Cloud sync is a switch on each project. A new project starts with it on while your plan has synced-project slots left.
When it is on, the project's contents are copied to Kosmico's cloud so collaborators and the web app can reach them. Secrets such as .env files and keys are never synced.
You can turn it off on projects you own: the cloud copy is removed, your files stay on the device, and anyone you shared it with loses access.

Cloud model

Your message, and whatever Kosmo reads to answer it, go through Kosmico to the cloud model you selected.
Conversations are stored so they sync across your devices, and you can delete them.
Kosmico does not use your documents or prompts to train AI models.

Local model

With a local Ollama model, nothing leaves your machine.

Research sources

arXiv · default, on requestSemantic Scholar · default, on requestOpenAlex · default, on requestPubMed · on request, added when relevant, subject-specificCrossRef · on request, added when relevantDBLP · added when relevantEurope PMC · added when relevant, subject-specific

Research pipeline

01You review the proposal: question, scope and sources
02Turn the question into research goals and search queries, with a quick web search for context
03Search the databases in parallel
04Merge duplicates by DOI, arXiv ID, PubMed ID or near-identical title
05Rate each candidate's relevance
06Search again, follow citations or look for gaps, one decision at a time
07Extract evidence from the full text of the most relevant papers
08Write the review section by section and build the Knowledge Graph

Stopping conditions

It finishes as covered only when it holds the minimum number of papers, its coverage score reaches the target, and every required research goal has supporting papers. The score is computed, not judged by a model: topics, methods, time span, source diversity, citation density and goals.
If Kosmo tries to finish earlier, it is sent back to search for the goals still open.
Otherwise it stops when the searches or passes allowed by the chosen depth run out, and the result is marked as completed with gaps, not as covered.

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.

Direct answers

Does Kosmico claim every specialist feature is stronger?

No. The claim being tested is workflow integration. Specialist tools may remain preferable for particular capabilities or established institutional processes.

How will the integration benefit be measured?

The benchmark measures completion, wall and human time, manual actions, application switches, artifact completeness, recovery, evidence quality, and cost across both real-world and controlled conditions.