Skip to content
CoSci

Introduction

CoSci is a local-first research co-scientist — a cockpit that runs the whole research cycle on your own machine and acts as a source-grounded scientific collaborator. You throw in a rough idea and it walks the full loop with you:

incubate idea → sharpen a specific, doable question → competing mechanisms → a killer experiment (with a hard kill-criteria gate) → record observations → update beliefs → write proposal / manuscript → wiki + canvas memory.

Everything runs on your machine. Your data stays local; every action is versioned and audited; sensitive actions are gated behind your explicit approval.

What makes it different

Most AI-for-science tools are cloud SaaS that stop at literature search or hypothesis generation. CoSci is built on the three things those tools trade away:

  • Local-first & private. Runs on-device, bring-your-own-keys, data never leaves. Your unpublished results and grant ideas stay yours.
  • Falsification-first. It models the discipline of the scientific method: competing mechanisms → a killer experiment that can’t be Ready without hard kill-criteria and named artifact risks.
  • Human-in-the-loop. Sensitive actions are enqueued, not executed. You stay in command.
  • Grounded, not generated. Literature and citation checks hit real databases through ToolUniverse, so citations are real and audited — it won’t invent a paper.

Under the hood

  • Backend: FastAPI + stdlib sqlite3 (no ORM). Serves on port 8799.
  • Frontend: React + Vite, served same-origin by the backend.
  • Offline & online: a built-in heuristic engine works with zero keys; add a real LLM and real scientific tools when you want them (online by default when available).
  • Open source: released under AGPL-3.0.

Where to next

  1. Install — one command gets the cockpit running.
  2. Quick start — your first idea → question → experiment.
  3. Providers & keys — wire up an LLM, or stay offline.
  4. Worked example — a full reproducible run.