01The Kartha Clinical Context Platform
Your product surfaces
Kartha Console / Explorer · your custom apps · Claude Desktop or any MCP client
Kartha Agent SDK + Accelerators
guardrail runtime · 50-skill clinical catalog · 10 packaged agents · Kartha Clinical AI Workflows (headless, cohort) · sessions & streaming · /v1 REST + SSE
Kartha MCP FHIR Server
fhir_read · fhir_search · fhir_create · fhir_update — any FHIR R4 resource · auth · tenancy · routing gate · summarizer · notes RAG
Your EMRs — FHIR R4
EPIC · Oracle Health / Cerner · OpenEMR · GCP Healthcare API · any FHIR R4 server
◆ Everything runs in your cloud. BYOC into your AWS or GCP account · model inference via your own provider · notes index in your Aurora pgvector · vendor auth quirks live in the provider layer, so your agents are EMR-agnostic. Any MCP client can also connect straight to the server — the SDK is there when you want guardrails, skills, and a runtime.
The safety posture, specifically
Ten invariants. Each one has an acceptance test, and the suite gates CI — if an invariant fails, the build doesn't ship.
02Your cloud. Your models. Your EMR data.
- AWS · GCP · AzureKartha deploys BYOC into your own account. Nothing runs on Kartha-operated infrastructure — you own the deployment, the keys, and the audit trail.
- Bedrock · Gemini Enterprise · Azure AI FoundryInference runs frontier models from your provider — AWS Bedrock, GCP Gemini Enterprise, Azure AI Foundry, or the Anthropic API. Your choice, your account, your rates.
- PHI never leavesThe patient chart stays inside your boundary — reached with your own EMR keys. Kartha carries the harness, never the data.
- EMR-agnostic by designVendor auth quirks live in the provider layer, so your agents are EMR-agnostic. Any MCP client can connect straight to the server — the SDK is there when you want guardrails, skills, and a runtime.
03Hospital-grade, by construction
- Security review, one lineThe whole platform deploys into your AWS or GCP account with your EMR keys. Kartha configures remotely — no PHI path, no runtime license server. “PHI never leaves our account.”
- A chart is 200K tokens. Your context window isn't.A deterministic, rule-based summarizer — no LLM in the compression path, no hallucination risk — plus per-code lab digests and prompt caching. Measured per answer by the console's token meter.
- Clinically vetted skillsExpert-reviewed workflows as content — SKILL.md, injected as context, never executed as code. Sepsis screening, med reconciliation, discharge planning, quality measures — or author your own in an afternoon.
- Private by constructionNotes embedded via your Bedrock into your Aurora pgvector. Forced row-level security, TTL'd derived cache, coverage receipts — and search degrades gracefully, never blocking a clinical question.
- Every answer carries its receiptsA collapsible tool trace — each request, the raw response, the guardrail badges that fired — plus cost, latency, and the token meter on every answer. Clinicians don't have to trust the agent was constrained. They can see it.
- A different layer than integration platformsRedox and Particle are integration-era — pipes between systems. Kartha is agent-era: the context, guardrails, and tools an AI agent needs to act on EMR data. Different layer, not a feature comparison.
04Going deeper?
The material your architects and your cloud security team will actually ask for — the architecture deep-dive, the security model, deployment runbooks — is provided on request, so it stays in a controlled channel. Tell us who you are and what you're building; the documents arrive by reply. Request the documentation →