For years, OpenCRAVAT has answered a deceptively simple question, what is known about this variant? by letting you assemble exactly the right annotators from our open app store and bringing hundreds of resources together in one place. The hard part was never the data. It was the synthesis: knowing which annotators to pick, how to weigh predictors that disagree, and how to turn a wall of columns into an interpretation you can act on.
Today we’re releasing something that takes that synthesis on directly. You can now simply ask OpenCRAVAT.
Meet OpenCRAVAT Chat (chat.opencravat.org) and the OpenCRAVAT MCP Server (https://mcp.opencravat.org/mcp) a conversational interface to the entire OpenCRAVAT annotation ecosystem.
💬 You: I found a KEAP1 variant of uncertain significance in a tumor. Should I be worried?
🧬 OpenCRAVAT: It’s a genuine VUS, but CHASMplus flags it as a likely driver (p = 0.037). Here’s exactly which evidence to trust, and what’s still missing. ↓
What it is
OpenCRAVAT Chat is an AI assistant that has OpenCRAVAT’s tools at its fingertips. Ask it a question in plain language “I found a KEAP1 variant of uncertain significance in a tumor. Should I be worried?” and it will look up the variant in OpenCRAVAT, pull annotations from across our resources, weigh the evidence, and hand you back a clear, structured interpretation.
It accepts variants however you have them: a protein change (KEAP1 R320Q), HGVS, genomic coordinates, a dbSNP rsID, or a ClinGen Allele Registry ID.
How it works and why you can trust it
Under the hood, OpenCRAVAT Chat uses the Model Context Protocol (MCP), an open standard that lets AI assistants call real software tools. We’ve exposed OpenCRAVAT’s core capabilities as MCP “tools”. The tools list available annotators, annotate a variant, convert a protein change to genomic coordinates, and more.
This matters for one big reason. The assistant doesn’t make up answers from memory, it retrieves them, live, from OpenCRAVAT. When it tells you a variant’s CHASMplus driver score, its population frequency in gnomAD, or whether CIViC has clinical evidence for it, those numbers come from actually running OpenCRAVAT, not from an AI’s recollection. And when the evidence isn’t there, it says so.
OpenCRAVAT Chat is built on the open-source LibreChat platform, connected to OpenCRAVAT through MCP. By default it runs on Anthropic’s Claude Sonnet 4.6, free of charge. If you have your own license, you can switch to any of the other popular chatbot models LibreChat supports.
OpenCRAVAT Chat is new, and there may be bugs/issues we have not identified. If you have problems, please report to us, so we can fix them quickly.
What it can do
A few things it does well:
It synthesizes. It doesn’t just dump scores. Ask about a variant where a dozen predictors disagree, and instead of averaging them into mush, it groups them by their calibrated ACMG/AMP evidence strength, explains why they split, and tells you what evidence would resolve the question. That’s the judgment call that usually takes an expert, made transparent.
It works in context. It reasons about the gene’s role, the protein domain a variant lands in, the tumor type, and co-occurring mutations, not a variant in isolation.
It scales from one variant to a whole case. Hand it a tumor’s mutation list and ask for a tumor-board-style summary, and it will rank the variants by actionability, surface the FDA-approved options, flag resistance signatures, note when a finding warrants germline testing, and pull it all into one coherent picture.
For example, hand it a lung-adenocarcinoma mutation profile such as EGFR L858R, TP53 R273H, KEAP1 R320Q, STK11 Q37*. You can ask for a tumor-board summary:
🫁 From a tumor’s mutation list to a tumor-board summary in one question
🥇 EGFR L858R → osimertinib (FDA-approved, first-line)
⚠️ STK11 + KEAP1 → immunotherapy-resistance signature, avoid checkpoint-inhibitor monotherapy
🧬 STK11 truncation → recommend germline testing
🧩 Four variants → one coherent, evidence-cited treatment strategy
Try it
- Just want to use it? Go to chat.opencravat.org and start typing.
- Want to plug OpenCRAVAT into your own AI workflow? Point any MCP-compatible client at the endpoint
https://mcp.opencravat.org/mcpand OpenCRAVAT becomes a tool your assistant can call, alongside the other tools you already use.
Not sure what to ask? Try one of these:
- “Is BRAF V600E a driver, and what therapies target it?”
- “My lung tumor has EGFR L858R and a new C797S, what does that mean for treatment?”
- “Here are 30 mutations from a tumor [list], which are likely drivers?”
- “How in general should I assess the impact of a non-coding variant?”
A note on responsible use
OpenCRAVAT Chat is a research and decision-support aid, not a clinical diagnostic. It’s grounded in OpenCRAVAT’s annotations, but AI assistants can still misread or omit things, so treat its output as a well-organized starting point, verify the underlying evidence, and leave clinical and germline interpretation to qualified professionals and your tumor board. As always, OpenCRAVAT is open and free.
What’s next
This is the first step toward something bigger: a future where OpenCRAVAT works together with other resources of the cancer-informatics ecosystem, including CIViC, cBioPortal, DGIdb, and more, through the same open protocol, so a single question can draw on all of them at once. We’re building toward that orchestrated, agent-driven interpretation.
Give it a try, tell us what you ask it, and let us know what it gets right, and where it can do better. Find us on GitHub, Bluesky, or the OpenCRAVAT community forum.
OpenCRAVAT is developed by the Karchin Lab at Johns Hopkins University with support from the NCI Informatics Technology for Cancer Research (ITCR) program.
