OpenMed Agent

The clinical agent that runs in your terminal — deterministic medical workflows, native coding and terminology tools, reviewable output.

OpenMed Agent gives clinicians, healthcare operators, and technical teams a terminal-native workspace for the work that sits between the chart and the claim: prior authorization and appeals, coding audit, clinical documentation, care coordination, and consumer health summaries.

[!NOTE]
OpenMed Agent is in preview. The product surface is real and usable today; the hosted medical-service tier is still evolving. See Preview status.

Install

# macOS / Linux
curl -fsSL "<install.sh URL shared during preview>" | bash

# Windows
irm "<install.ps1 URL shared during preview>" | iex

Install URLs are shared with evaluators during preview — request access.

Quickstart

openmed login   # Authenticate with OAuth
openmed         # Launch the agent
openmed --help  # Full CLI reference

Or bring your own key:

openmed config provider-set openai --api-key "sk-..."
openmed config provider-set anthropic --api-key "sk-ant-..."

Runs on GPT-5.5 / GPT-5.6 and Claude models. gpt-5.6-terra at medium reasoning is the default; switch model, reasoning effort, skill, and agent mode from the TUI or with openmed agent.

What it does

OpenMed Agent combines LLM reasoning with deterministic workflows and native medical tools. Describe the task in natural language, and the agent can:

• Review prior authorization and appeal cases against structured criteria
• Audit ICD-10 coding with HCC and RAF context
• Explain EOB and claims data in plain language
• Extract entities or de-identify clinical text
• Summarize consumer health records from Apple Health, Health Connect export, C-CDA, FHIR export, and labs files
• Triage inbox threads, draft reviewer-safe replies, and generate discharge handoffs
• Search PubMed and use protected terminology services for ICD-10, CPT, SNOMED, LOINC, RxNorm, MedlinePlus, HCC, and RAF

What ships in preview

• 62 native tools
• 13 deterministic workflows with draft/finalize
• 13 built-in skills
• 4 agent modes: clinical, consumer, coordination, plan
• 104 demo scenarios covering every capability area

Key capabilities

Protected medical services

• Clinical extraction — entity extraction, PII detection, and de-identification through protected service endpoints
• Terminology and coding — ICD-10, CPT, SNOMED, LOINC, RxNorm, MedlinePlus, HCC, and RAF, with validation, crosswalks, and PubMed-backed lookup
• Configurable deployment boundary — the operator runtime stays local while medical-service endpoints can be moved across hosted, cloud, or customer-managed environments

Clinical and operational workflows

• Prior authorization and appeals — review requests against deterministic criteria and structured evidence
• Coding audit — specificity review, compliance flags, HCC mapping, and RAF impact
• Claims explanation — patient-friendly EOB and billing explanations with clear next steps
• Clinical documentation — structured SOAP-style documentation from notes or transcripts
• Care coordination — inbox triage, reviewer-safe patient drafts, discharge and PCP handoffs, follow-up tasks
• Consumer health — record normalization, timeline and trend analysis, visit-prep questions, narratives, reconciliation, optional education topics, and optional FHIR output

Agent runtime

• Project instructions — drop an OPENMED.md in your project root to shape agent behavior per workspace
• Permission policy — rule-based auto / acceptall / denyall / plan modes, with per-tool allow, deny, and ask rules from project or user settings
• Tool safety classification — every tool carries is_read_only, is_concurrent_safe, and is_destructive metadata used by the permission system
• Oversized result handling — large tool outputs persist to disk automatically, leaving compact stubs in context
• Runtime diagnostics — /config shows effective settings, loaded sources, and active project instructions

Agent experience

• Interactive TUI — sessions, themes, model and skill switching, and workflow execution in the terminal
• Draft and finalize — reviewable cards, artifacts, provenance, and workflow diffs
• Skills — built-in clinical skills that shape how the agent approaches domain work
• Plan auto-advance — structured plans with deterministic, tool-based progress tracking
• Session persistence — save, restore, fork, and roll back conversations
• Optional MCP — connect external medical or institutional systems without changing the native tool surface
• Self-update — built-in binary updates with release checks

Preview status

• The product surface is real and usable today
• The hosted medical-service tier is still evolving
• Some service-backed capabilities are operated by OpenMed during preview rather than fully self-serve
• Workflow, deployment, and integration details may continue to tighten as the product hardens

During preview, OpenMed operates the protected clinical-service endpoints so evaluators do not have to deploy extraction and terminology infrastructure themselves.

Where the boundary sits

The operator runtime — sessions, artifacts, project instructions, and permission rules — stays on your machine. Medical capabilities call configured protected endpoints when those paths are invoked, and those endpoints can be moved across hosted, cloud, or customer-managed environments.

Three things worth stating plainly:

• Not fully offline. Invoking a protected capability means a call to a configured endpoint.
• PHI handling modes are settings, not guarantees. They are operator-visible workflow controls, not a blanket promise that every code path enforces the same policy.
• No telemetry. No built-in analytics or phone-home tracking ships with the product.

Reviewability is a core product characteristic: final clinical artifacts and care-coordination outputs are designed to be inspected before use.

Website · Documentation · X/Twitter · LinkedIn