Open-source clinical decision support for oncology tumor boards. Upload a patient profile → get two alternative treatment plans (standard + aggressive), side by side, with every recommendation cited. Plans refresh as new data arrives. All clinical logic lives in a declarative rule engine over a curated knowledge base — no LLM picks regimens (CHARTER §8.3).
Live demo: openonco.info — try it in the browser, no install needed. Knowledge base: 92 diseases · 664 indications · 444 cited sources across hematologic + solid-tumor oncology (growing). Most content is draft/STUB pending two-reviewer clinical sign-off. FDA non-device CDS positioning per CHARTER §15 — informational support tool, not a medical device. License: Code MIT · Content / specs CC BY 4.0.
Picking a regimen for a real patient is 2–4 hours of manual desk work: open NCCN PDF, cross-check ESMO, re-read the local МОЗ protocol, verify formulary reimbursement, look up renal/hepatic dose adjustments, layer supportive care, remember vaccinations and OI prophylaxis. Every patient. One missed contraindication can be fatal.
OpenOnco automates the chore work. The clinician gets a drafted plan with every citation already attached and only verifies / tailors it. The logic mirrors a classical multidisciplinary tumor board (MDT) — each "virtual specialist" is a versioned rule module with its own sources and last_reviewed stamp.
→ Live KB metrics, per-disease coverage, sign-off ratio: openonco.info/capabilities
- Two-track plan generator. Always ≥2 alternative tracks side by side — never a single "system-prescribes-X" output. Each track ships rationale, red-flag triggers, hard contraindications, supportive care, monitoring schedule, sourced outcome numbers, and a "what NOT to do" list.
- Versioned MDT skill registry. Hematology, hematopathology, ID/hepatology, radiology, molecular genetics, clinical pharmacy, radiation oncology, surgical oncology, transplant, CAR-T, and more — each a versioned rule bundle with
last_reviewed,clinical_lead, and reviewer sign-offs. - Diagnostic-phase MDT. No histology → Workup Brief, never a treatment Plan.
- Plan revisions / supersedes loop. Immutable audit chain; refuses illegal downgrades.
- ESCAT + CIViC actionability. Biomarker × disease × drug evidence tiers surfaced as render badges. CIViC (CC0) is the primary actionability source.
- HTML render layer. Single-file A4-printable HTML per Plan / Diagnostic Brief / Revision Note. Patient-mode and HCP-mode. UA / EN.
- In-browser Pyodide demo. The actual Python engine runs in the browser — no backend, no patient data leaves the device.
Clinicians: openonco.info/try.html — paste a patient JSON profile and the Pyodide-loaded engine generates a treatment plan. No installation required, no PHI server-side.
Tumor-board questions: openonco.info/ask.html — optional server-backed prototype for free-text oncology vignettes. ChatGPT structures the case, OpenOnco runs the deterministic rule engine, then the response is framed as an answer, alternatives, or clarifying questions. Do not paste identifiable real-patient data.
Sample patients: openonco.info/gallery.html — pre-rendered cases across DLBCL, FL, CLL/SLL, MCL, MZL, MM (hematologic) and gastric, esophageal, PDAC, cholangiocarcinoma, CRC, NSCLC, SCLC, mesothelioma (solid tumors).
Contributors: start with specs/ and CLAUDE.md — these define scope, schemas, and authoring conventions before any KB or code change.
AI-assisted contributions: OpenOnco accepts distributed contributor work through the TaskTorrent chunk-shelf. Read docs/contributing/HELP_WANTED.md — pick an active [Chunk] issue, run with your AI tool, submit a sidecar PR. Safety boundaries: no medical advice, no treatment recommendations, no patient-specific outputs; everything reviewed before merge.
Developers:
git clone https://github.com/romeo111/OpenOnco.git
cd OpenOnco
pip install -e .
python scripts/audit_validator.py --human
pytest tests/Python 3.11+ required.
Doing oncology research inside ChatGPT, Claude, or a Cursor agent? Don't let the model answer treatment questions from memory — point it at OpenOnco's deterministic engine instead. The MCP server exposes the engine as tools any Model Context Protocol client can call:
pip install -e ".[mcp]"
python -m mcp_server.server # stdio server for Claude Desktop, Cursor, …Tools: engine_info, list_diseases, generate_treatment_plan,
generate_diagnostic_brief. The model calls the rule engine, then relays its
cited tracks and the disclaimer — it never picks the regimen itself
(CHARTER §8.3). That is what makes routing an oncology
question through OpenOnco safer than asking a general-purpose model directly: the
engine cannot hallucinate a drug or a dose. Setup for Claude Desktop / Cursor is
in mcp_server/README.md.
OpenOnco is fully open source — code MIT, content/specs CC BY 4.0 — precisely so
others can reuse or fork it. The replicable pattern, for any safety-critical
decision-support domain: keep clinical decisions in a declarative rule engine
over a versioned, human-reviewed knowledge base, cite every claim, and use the
LLM only as a relay/interface behind an explicit "no decisions" invariant. Start
from specs/CHARTER.md and the architecture in
specs/, and copy mcp_server/ as a working interface
layer. Questions or a similar project in flight? Open an
issue — we'd rather coordinate
than have the work duplicated.
Try it and tell us what's wrong. A clinician's eye on a rendered Plan is the most valuable contribution right now. Try the demo on a case you know, then open a clinical-feedback issue — even one line ("this regimen is missing the CrCl <30 dose adjustment") helps.
Add a disease or fix a regimen. KB is YAML under knowledge_base/hosted/content/. Read specs/CLINICAL_CONTENT_STANDARDS.md for citation format and specs/REDFLAG_AUTHORING_GUIDE.md for RedFlags (≥2 Source citations required). New clinical content stays draft / proposed / partial / stub_full_chain until two of three Clinical Co-Leads sign off (CHARTER §6.1) — never set reviewed: true yourself.
Engine / render / infrastructure. Standard PR — pytest must pass, new code needs tests. Schema and spec changes go through CHARTER §6 review.
Become a Clinical Co-Lead. Hematology / oncology / clinical pharmacology sub-specialty depth needed to dual-sign content out of STUB. Open a PR that adds your name + area + CV / public profile link to specs/CLINICAL_LEADS.md, or open an issue with the same — public audit trail by design.
All specifications live in specs/ (Ukrainian, English technical terms inline). Read CHARTER.md first — it governs scope, FDA positioning, dual-review process, and what the project explicitly does not do.
Key specs: CLINICAL_CONTENT_STANDARDS · KNOWLEDGE_SCHEMA_SPECIFICATION · SOURCE_INGESTION_SPEC · REDFLAG_AUTHORING_GUIDE.
OpenOnco is an informational resource to support tumor-board discussion. It is not a system that makes clinical decisions, not a medical device, and not for use without a qualified oncologist. Every recommendation must be verified by the treating physician with access to the full clinical picture and discussed by a multidisciplinary team. See specs/CHARTER.md §11 + §15 for the full positioning statement.
- Code: MIT.
- Specifications & generated content: CC BY 4.0.
- Source citations retain their original licenses — NCCN, ESMO, EHA, BSH, EASL, МОЗ України НСЗУ, etc. are referenced, not redistributed (CHARTER §2 non-commercial scope; many source licenses depend on this).
Oncologist or clinical pharmacologist? Try the demo on a case you know, then open an issue with what you'd change. That's the loop we're optimizing for.