Eyes on the patient, not the keyboard — the visit documents itself.

CoDoc is an ambient clinical copilot. It listens to the doctor–patient conversation, understands it in real time, and turns it into a real, standards-compliant medical record — with zero typing.

Built at the YC × Medplum Agentic Healthcare Hackathon (Aug 2026) on Medplum, Deepgram, and Claude.

• Listens — Deepgram Nova-3 streaming with speaker diarization and medical keyterm boosting
• Understands — Claude infers who the clinician is, extracts symptoms (with the patient's verbatim words), vitals, allergies, meds, and intake details (DOB, phone, address, pharmacy)
• Registers patients by voice — say your name: returning patients get their full history pulled up; new patients get a chart created on the spot
• Catches what humans miss — cross-checks the live conversation against the stored record (e.g. flags an amoxicillin order against a documented penicillin allergy)
• Acts on command — "CoDoc, refer her to orthopedics" → drafts the order note and queues a real FHIR ServiceRequest + Task for the specialist; prescriptions become MedicationRequests
• Closes the visit — one click signs a SOAP note into the chart as a DocumentReference
• Three-sided product — PCP runs visits, the specialist works a referral queue (accept → patient appears in their panel), and the patient sees their own record, conversations, and referral status in plain language

Everything is written to Medplum as conformant FHIR (11+ resource types), live, as the conversation happens.

Prereqs: Node 22.5+ (uses built-in node:sqlite), npm, and a microphone.

git clone https://github.com/Nehanth/CoDoc.git
cd CoDoc
npm install

Create .env.local in the project root:

# Deepgram — https://console.deepgram.com (free credits)
DEEPGRAM_API_KEY=your_deepgram_key

# Anthropic — https://console.anthropic.com
ANTHROPIC_API_KEY=your_anthropic_key

# Medplum — https://app.medplum.com → Project → Clients → create client
MEDPLUM_BASE_URL=https://api.medplum.com/
MEDPLUM_CLIENT_ID=your_medplum_client_id
MEDPLUM_CLIENT_SECRET=your_medplum_client_secret

Then:

npm run dev

Open http://localhost:3000 (must be localhost or HTTPS — the mic requires a secure context).

Patient accounts self-register on first login for any patient that exists in your Medplum project; the password you type first becomes theirs.

1. Sign in as pcp, hit Start session
2. Clinician: "Hi, can I get your name and date of birth?" — Patient: "I'm Emily Carter, March 14th, 1998…" (phone, address, pharmacy, penicillin allergy, meds)
3. Patient: "I've had pain in my right knee for two months, since I started training for a 10K" → watch the chart bloom: symptoms with quotes, body map, vitals as spoken
4. Clinician: "I'll start you on amoxicillin for that sinus thing" → 🚨 contraindication card (penicillin allergy)
5. Clinician: "CoDoc, refer her to orthopedics" → agent drafts the note, order lands in the queue
6. End visit · sign note → SOAP note signs itself
7. Log out → specialist → accept the referral (patient's phone is right there) → Emily appears in their Patients panel
8. Log out → Emily Carter → she sees her record, the whole conversation, and "Accepted — expect a call to schedule"

Check app.medplum.com afterwards — every resource is real FHIR.

• Login is demo-grade (plaintext SQLite, localStorage session) — hackathon scope, not production auth
• Pharmacy integration is a labeled dummy; referral receiving is intentionally out of scope (any FHIR-native system can work the Task queue)
• codoc.db (accounts) is created on first run and gitignored