CoSci
🔒
No cloud. No exceptions.
CoSci runs on your machine. Your unpublished data, hypotheses and grant ideas never leave it — online mode sends only what you approve, only to the provider whose key you supplied.
Local-firstFalsification-gatedHuman-in-the-loop Real, cited toolsWhole lifecycleBYO-keys
Your whole research cycle,
on your own machine.
CoSci is an open-source research co-scientist that runs the full research cycle locally — idea to manuscript — grounded in real scientific tools, gated by the scientific method, and always under your approval.
Runs offline & online · your data never leaves · open source under AGPL-3.0
The wedge
The co-scientist that doesn't take your data — or make things up.
Cloud AI-for-science trades away privacy, rigor and control. CoSci is built on exactly those three.
📚
It won't invent a citation.
Every literature sweep and citation check hits real databases through ToolUniverse. Claims are audited against references that actually exist — grounded, not generated.
✋
It won't act without you.
Sensitive actions are enqueued, not executed. Three task classes — needs input, needs approval, AI-can-execute — keep you in command of every step.
🎯
It forces the killer experiment.
An experiment can't be marked Ready until it has hard kill-criteria and named artifact risks. The discipline of the scientific method, encoded as a gate.
🧠
It remembers the whole story.
Idea → question → mechanism → experiment → belief update → manuscript, all versioned and auditable. Your research memory compounds instead of scattering across chats.
⚡
Works offline, better online.
A built-in heuristic engine runs with zero keys. Add an LLM and real tools when you want them; CoSci uses the best provider available and falls back gracefully.
The lifecycle
One cockpit for the entire research loop.
Not a point tool — CoSci walks every stage with you, and versions all of it.
1
Incubate
Throw in a rough idea; get ranked candidate questions.
2
Sharpen
Narrow to one specific, doable question worth answering.
3
Mechanisms
Lay out competing explanations, not a single guess.
4
Killer experiment
Design the test — gated by hard kill-criteria.
5
Observe
Record real observations against your predictions.
6
Update beliefs
Revise confidence as evidence comes in.
7
Write
3-agent proposal & manuscript packets, with a critic.
8
Remember
Everything lands in a versioned wiki + canvas memory.
The hinge of the whole thing: an experiment can't be Ready without hard kill-criteria and named artifact risks.
Grounded in real science
Live scientific databases — not a language model's memory.
Literature sweeps, citation checks and protein/structure context reach real sources through the ToolUniverse layer. Real and cited.
PubMed · biomedical literatureEurope PMC · open full-textSemantic Scholar · citations & graphUniProt · proteinsInterPro · domainsPDB · structures
A complete worked example
DNA replication–transcription conflict, end to end.
One reproducible run in the repo does the whole loop with real tools:
- Incubates the idea and sweeps the literature
- Resolves 4 E. coli proteins (RpoB, DnaB, Mfd, Rep) to real UniProt accessions with InterPro domains + PDB structure counts
- Designs a killer experiment that passes the kill-criteria gate
- Writes a proposal, then citation-checks the claims against real 2024–26 papers
$ uv run python walkthrough_conflict.py incubate idea ................. ok literature sweep .............. ok real resolve proteins (UniProt) .... ok RpoB DnaB Mfd Rep domains + PDB (InterPro/PDB) .. ok killer experiment ............. ok kill-gate passed proposal + citation gate ...... ok 2024–26 refs → test-results/conflict-walkthrough/REPORT.md
How it works
Running in under five minutes.
One machine, your keys (or none), the whole workflow.
01
Install — one command
Double-click start.command (or run it). First run builds the UI, seeds a profile, and opens the cockpit at localhost:8799.
./start.command
02
Bring your keys — or don't
Point it at the Claude or Codex CLI, Gemini, or DeepSeek. No key? The heuristic engine runs fully offline.
RC_OFFLINE=0 # or 1 for hard offline
03
Work the cycle
Capture an idea, incubate it, sweep the literature, build mechanism context, design the killer experiment — with you in the loop.
open http://localhost:8799
Free & open source — and looking for beta users.
No price, no trial, no bot in your lab. CoSci will be released under AGPL-3.0 — local-first, bring-your-own-key, or fully offline. The private beta is open now.
macOS / Linux · runs on uv + Node · see how it works in the
install guide.
FAQ
Is my research data private?
Yes. CoSci is local-first: it runs on your machine and stores everything in a local SQLite database. Nothing is uploaded anywhere by default. In online mode, only the content you approve is sent — and only to the AI provider whose API key you configured.
Can it hallucinate citations?
It's designed not to. Literature sweeps and citation checks go through ToolUniverse to real databases (PubMed, Europe PMC, Semantic Scholar, UniProt, InterPro, PDB). A citation-existence gate audits each claim against references that actually exist, so 'it won't invent a paper' is a concrete promise, not a hope.
Do I need an API key?
No — a built-in heuristic engine runs the whole workflow offline with no keys at all. Add a provider (the local Claude/Codex CLI, Gemini, or DeepSeek) when you want stronger reasoning and real tool calls. CoSci picks the first available provider and falls back automatically.
What can it actually do, end to end?
It walks the full research cycle: incubate an idea into ranked questions, sharpen one, lay out competing mechanisms, design a killer experiment behind a kill-criteria gate, record observations, update beliefs, and draft proposal/manuscript packets — all versioned in a wiki + canvas memory. A reproducible worked example (a DNA replication–transcription conflict) ships in the repo.
Who is it for?
Working scientists — grad students, postdocs and PIs — who want an AI collaborator they can trust with unpublished work. It's especially aimed at life sciences, but the workflow is domain-general.
Is it open source?
Yes — CoSci is released under AGPL-3.0. You can read it, build it, and self-host it. It's an actively developed project provided as-is.
Is there a hosted or team version?
Not yet. The local-first app is the product today. Managed cloud sync and a shared lab/team workspace are on the roadmap — the same 'your data stays yours' story, for people who'd rather not self-host.