An open multi-agent research partner that turns a research goal into ranked, literature-grounded hypotheses.
Describe what you want to understand, such as a mechanism, a drug-repurposing question or an unexplained observation. Open Co-Scientist searches the literature, drafts competing hypotheses, critiques and fact-checks them, runs an Elo tournament of scientific debates, and evolves the strongest ideas. You get a report with ranked hypotheses, the mechanism behind each, a proposed experiment, and the sources every claim rests on. You can steer the run while it works.
It is an open, independent implementation of the agent design in Google's AI co-scientist. It is not affiliated with or endorsed by Google.
An illustrative Express workflow with scripted responses and sample results, shown at double speed.
Adapted in the Open Co-Scientist palette from Figure 2 of Gottweis et al., Towards an AI co-scientist, page 9 (2025). Original diagram.
The supervisor model advises on each next step, and a deterministic policy enforces budgets, forced transitions and convergence from live statistics such as pool growth, Elo stability and match coverage. Every step is a durable task, so a run survives restarts and resumes where it stopped.
| Capability | What you get |
|---|---|
| Cooperating agents | Six agents under a supervisor: generation, reflection, ranking, evolution, proximity and meta-review, following the published design. |
| Evidence you can check | Hypotheses cite retrieved sources with [C*] keys, and retracted papers are flagged. For the ranked finalists featured in the report, each claim is checked against its passages for support or contradiction. |
| Literature and database tools | A bundled MCP server provides PubMed and PMC full text, OpenAlex, Europe PMC, arXiv and preprints, ChEMBL, UniProt, STRING, Reactome, Open Targets, Ensembl, gnomAD, the GWAS Catalog, ClinicalTrials.gov, and web search when a search key is configured. |
| Steering and questions | Add constraints or ask about the evidence while a run works, and attach your own documents. |
| Safety screening | Screening of the goal, each hypothesis and the final report. |
| Bring your own key | Use your preferred provider and larger run sizes; keys are encrypted at rest. |
| Deterministic test backend | Tests and browser suites exercise the whole pipeline without an API key using COSCIENTIST_TEST_DOUBLE=deterministic. |
- Free model routes. The hosted service runs on free OpenRouter routes: Ling 3.1 Flash, with free Nemotron fallbacks. They are slower and weaker than frontier models, their zero price is the provider's decision and can end, and results depend heavily on the model.
- Capacity. Without your own key, each browser gets 11 free credits a day by default (Express 2, Standard 3, Extended 5, Ultra 7), behind per-network and global daily caps. The free routes are rate-limited for the whole deployment, so a busy day can park runs until the daily reset. Runs take from tens of minutes to several hours, depending on size and provider speed.
- One server. The API is a single process on SQLite. It favors durability and simplicity over scale.
- Hypotheses, not findings. Passing tests shows the software works, not that its hypotheses are right. Treat every output as a starting point for expert review and experiment. The evaluations say what has and has not been measured.
- Free routes and your data. Free model routes may keep the prompts they receive. Do not submit confidential or unpublished work without your own key and a provider whose terms you accept.
- Private by obscurity. Runs belong to a per-browser ID, not a verified account.
You need Python 3.12, Node.js 22.13+ and Bun 1.3.14 on macOS or Linux (on Windows, use WSL2).
git clone https://github.com/guy915/open-co-scientist.git
cd Open-Co-Scientist
make setup # Python virtual environment and locked frontend dependencies
make start # API on :8008, workbench on :5173, MCP server on :8888make start opens localhost:5173 when it is ready,
and stops anything already listening on those three ports. A run needs a
model credential: without one, requests fail with "No model is available right
now". Set OPENROUTER_API_KEY in the root .env that make setup created; see
.env.example and the
full API configuration. The MCP server reads its own
engine/mcp_server/.env (template).
make check runs the CI checks apart from the image builds: lint, types,
every test suite, the evaluation smoke test, the production build and the
browser tests. make help lists
every target, and local setup covers details and
troubleshooting.
flowchart TB
ui[React workbench] -->|HTTP requests| api[FastAPI API]
api -->|Progress via server-sent events| ui
subgraph engine[Python package: co_scientist]
api -->|Enqueue tasks| runtime[Durable task runtime]
runtime --> agents[Scientific agents]
agents --> models[Model gateway]
agents --> retrieval[Retrieval]
api --> store[(SQLite store)]
runtime --> store
end
models --> providers[Model providers]
retrieval --> mcp[MCP server]
mcp --> sources[Literature and databases]
classDef interface fill:#dcfce7,stroke:#15803d,color:#14532d
classDef service fill:#dbeafe,stroke:#2563eb,color:#1e3a8a
classDef external fill:#f3e8ff,stroke:#9333ea,color:#581c87
class ui interface
class api,runtime,agents,models,retrieval,store service
class providers,mcp,sources external
Open Co-Scientist is one Python package, co_scientist
(engine/src/co_scientist/), served by FastAPI. It is layered from the
composition root through the API, orchestration, the scientific agents and
domain logic down to the platform (models, retrieval, sandbox and the SQLite
store) and core, and import contracts enforce that layering in CI. A React
workbench (app/frontend/) streams progress over server-sent events, and a
reference MCP server (engine/mcp_server/) provides the literature and
database tools. Read the architecture, the
engine architecture and the
documentation index.
Scientists and developers are both welcome. You can report a bug, propose a research question or an evaluation dataset, reproduce a result, or send code. Start with CONTRIBUTING.md, and ask questions in Discussions. Everyone follows the Code of Conduct. Report vulnerabilities privately as described in SECURITY.md.
If Open Co-Scientist contributes to your work, cite the archived v1.0.0 release: Guy Barel, Open Co-Scientist (version 1.0.0), Zenodo, 2026, doi:10.5281/zenodo.23366148. CITATION.cff holds the same metadata, and GitHub's Cite this repository menu exports it as APA or BibTeX.
First-party software is licensed under Apache 2.0. NOTICE records third-party software and data; vendored material, cited papers and scientific data keep their own terms.
This project builds on Google's AI co-scientist research:
- Gottweis et al., Towards an AI co-scientist, 2025.
- Accelerating scientific discovery with Co-Scientist, Nature, 2026.
Much of the web app's interface is modelled on Google's Idea Generation agent in Google Agentspace, now the Co-Scientist agent in Gemini Enterprise. The layout, flow and visual style follow that product; the code is independent work.
It also draws on Science Skills
(vendored under vendor/), Jataware's open-coscientist,
Sakana AI's AI Scientist and
Pi Agent.

