Last EHR

Status: early / alpha. Use synthetic data only. License: Apache-2.0.

Works with Epic and Oracle Health sandboxes

Last EHR can be the approval gate in front of the Epic FHIR sandbox and the Oracle Health secure sandbox. The human sees the Observation those servers accepted (Epic flowsheet code plus encounter, Oracle legal-authenticator performer and an allowlisted vital), and the agent gets a read-back diff after the save. Synthetic sandbox patients only. This is not a clinical or compliance claim. The adapters stay behind LASTEHR_EXPERIMENTAL_FHIR_ADAPTERS=1 because those sandboxes do not support delete: an approved write persists, and the safety eval cannot clean up. Ten-minute setup is in docs/adapters/epic-cerner.md.

export LASTEHR_EXPERIMENTAL_FHIR_ADAPTERS=1
export FHIR_BACKEND=epic
export LASTEHR_CLIENT_ID='your-non-production-client-id'
export LASTEHR_PRIVATE_KEY_PATH="$PWD/lastehr-sandbox.pem"
export LASTEHR_KEY_ID='lastehr-epic-1'
export LASTEHR_ENCOUNTER='Encounter/eMT95RPaM-HxLc3NeIA5qlQ3'

Then ask: Record a heart rate of 72 bpm for patient erXuFYUfucBZaryVksYEcMg3. The card shows flowsheet code 8 and LOINC 8867-4 before anything is saved. Oracle uses FHIR_BACKEND=oracle-health, LASTEHR_SCOPES, and an oral temperature (8331-1) with LASTEHR_PERFORMER.

The same sandboxes are wired into @lastehr/mcp 0.5.0 in this checkout. That version is prepared and not published; npm latest below is still the last published package. From a checkout:

npx tsx packages/mcp/src/cli.ts init --backend epic
npx tsx scripts/sandbox-proof

init --backend epic or oracle prints placeholder env (LASTEHR_BACKEND, LASTEHR_CLIENT_ID, LASTEHR_PRIVATE_KEY_PATH, LASTEHR_KEY_ID, LASTEHR_JWKS_URL, and LASTEHR_ENCOUNTER or LASTEHR_PERFORMER plus LASTEHR_SCOPES for Oracle) and the setup steps. doctor checks the key, JWKS kid, token exchange, granted scopes, and search scope without printing secrets.

Start here

The first step is the published package. It does not create an account.

npx -y @lastehr/mcp init

npm latest is @lastehr/[email protected]. The JSON this command prints must list @lastehr/[email protected] in args. If it names another version, stop and run npx -y @lastehr/[email protected] init instead. This page describes that release. init in 0.4.1 accepts --client json (the default), claude-code, or cursor. It prints a client configuration and exits. It does not start a server, register the client, or contact a FHIR endpoint, and it has no flag that selects a hosted demo.

Default output:

{
  "mcpServers": {
    "lastehr": {
      "command": "npx",
      "args": ["-y", "@lastehr/[email protected]"],
      "env": {
        "MEDPLUM_ACCESS_TOKEN": "<replace-with-a-least-privilege-token>"
      }
    }
  }
}

--client cursor adds "type": "stdio". --client claude-code prints claude mcp add lastehr -- npx -y @lastehr/[email protected] plus comments. The generated command is pinned to the version that ran init.

No-account synthetic backend

The no-account backend 0.4.1 accepts is this repository's local HAPI stack. Set FHIR_BACKEND=hapi and FHIR_BASE_URL or HAPI_BASE_URL. No Medplum token. The stack has no auth, Compose binds it to loopback, and it is for synthetic records only.

Node 22.18+ (or 24.2+) and a running Docker daemon:

git clone https://github.com/cbetz/last-ehr.git
cd last-ehr
npm install
npm run demo:local

demo:local starts HAPI and Postgres, seeds synthetic patients, and opens the fixed walkthrough at http://localhost:3000/demo. It does not create or read .env.local. The first image pull can take several minutes. Press Ctrl-C to stop Next.js. npm run demo:local:down removes the stack.

In the configuration init printed, replace the env block with:

"env": {
  "FHIR_BACKEND": "hapi",
  "FHIR_BASE_URL": "http://localhost:8080/fhir"
}

Keep the rest, including "type": "stdio" when --client cursor printed it. Save it where the client reads MCP config (for Cursor, ~/.cursor/mcp.json or .cursor/mcp.json). With FHIR_BACKEND=hapi, 0.4.1 does not read MEDPLUM_*.

The Claude Code line does not carry those variables. Export them in the shell that launches Claude Code:

export FHIR_BACKEND=hapi
export FHIR_BASE_URL=http://localhost:8080/fhir

After HAPI is up, this check prints no patient records:

FHIR_BACKEND=hapi FHIR_BASE_URL=http://localhost:8080/fhir npx -y @lastehr/mcp doctor --connect

In 0.4.1, doctor --connect checks FHIR R4 metadata and a count-only Patient search. It does not check other resource permissions, writes, or a client's approval UI.

npm run mcp:demo is a separate checkout-only lab, outside the published @lastehr/[email protected] package. It exposes search_patients and show_patient_info for four fixture patients. See the MCP guide.

Local stack failures: Troubleshooting the local lab.

Other ways in

These are optional. The published MCP package and the local synthetic stack above are the first path.

Goal	Start here
Test an MCP (stdio) implementation of the write protocol	npx @lastehr/agent-write-conformance — conformance guide
Have an assistant draft synthetic work, then review it separately	npm run work:demo — Assistant Work Lab. Add --agent --model-env .env.local only when a model key is already configured.
Connect an external MCP assistant to durable synthetic work	Work MCP connector: npm run --silent work:mcp -- --print-config
Deploy the web app	Quickstart: Vercel, with your own Medplum project and model key
Run the web app locally	Quickstart: Medplum, or npm run demo:local for the fixed synthetic walkthrough
Open the hosted synthetic demo	https://www.lastehr.com/demo — no sign-up. Ask Find patients named Smith, then Record a heart rate of 72 bpm for Maria Garcia, and inspect the approval card before anything is saved.
Read the protocol	Approval-Gated Agent Writes on FHIR, v0.1 draft
See the FHIR read surface	FHIR coverage
Inspect two read-only tools on four fixture patients	npm run mcp:demo -- --client claude-code from a checkout. Not the published package.

What it is

Read the chart broadly. Write only what a human approved. Last EHR is the agent layer for a headless FHIR EHR. It does three things, and the write protocol is only one of them.

It reads the chart broadly. 25 of US Core 9.0.0's 27 readable resource types, across 23 patient-scoped sections, following references so authors, encounters, and provenance stop being dead pointers. Counted in the open, with the ceiling published rather than implied, in FHIR coverage.

It is built so the agent cannot fake an absence. "She has never had a flu shot" is the answer a chart agent must never invent, so a read reports when its window was capped, when a reference lookup was refused, and when a code filter matched nothing in a section that does hold records. Filters that a section cannot apply are refused with the legal values instead of silently dropped. These are safety-boundary tests, not best-effort behavior.

Every write is a proposal. Show the exact resource, require an explicit human decision, commit exactly what was reviewed, and record what created it. That half is a small protocol, Approval-Gated Agent Writes on FHIR (v0.1 draft): Proposal → Decision → Commit → Audit. Two implementations run here, a web approval card and MCP elicitation-gated write tools, plus a conformance suite that checks implementations with its own FHIR reads. Independent implementations and criticism are invited.

Five FHIR backends sit behind one interface. This repository is the reference implementation: bring your own FHIR backend and model key for a real agent (Medplum for the authenticated path, or any verified adapter for synthetic evaluation), or start with the zero-key local synthetic walkthrough.

Last EHR is a layer, not an EHR. It runs on top of a headless FHIR backend (Medplum for authenticated use; HAPI FHIR, Firely Server, Aidbox, or Oystehr for synthetic evaluation) and talks to it over the FHIR API. It is not the system of record and never bundles or forks the backend. The web agent has no separate chart database, but processes chart data and sends chart context to the configured model provider. The opt-in local Work Lab persists synthetic requests, run progress, proposals, and decisions. See the HIPAA deployment groundwork before planning a clinical deployment.

Status: early / alpha. APIs, structure, and scope will change. Use synthetic data only. · License: Apache-2.0

Try the live demo: no sign-up, synthetic data, and every write goes through the approval gate.

What it does

• A chat agent (Vercel AI SDK) with FHIR tools. It reads the chart (search patients, view a patient chart, read one filtered chart section) and writes to it (add a note, record an observation, create a follow-up task), streamed and rendered as structured cards.
• Writes are confirmation-gated: the agent proposes a write, you approve it, and only then is it saved. Nothing touches the chart without your click. Approved writes carry the standard AIAST security label, with optional Provenance recording author-agent and human-verifier roles (role labels, not user identity).
• The gate is a specified protocol, not just app behavior: Approval-Gated Agent Writes on FHIR (v0.1 draft) names the rules — only an explicit approval commits, every ambiguous outcome fails closed, policy can only narrow the gate — and the conformance suite tests an implementing MCP stdio server's gate mechanics with independent FHIR reads. CI runs it against this repository's own server on every pull request and merge.
• Authentication, multi-tenancy, and access control are delegated to your Medplum project (Project = tenant, ProjectMembership = user, AccessPolicy = RBAC). Last EHR doesn't reimplement any of that, and writes are bounded by your AccessPolicy.

Support status

Last EHR is Medplum-supported for authenticated deployments and includes a local, no-auth HAPI FHIR mode for synthetic-data evaluation. Firely Server (FHIR_BACKEND=firely), Aidbox (FHIR_BACKEND=aidbox), and Oystehr (FHIR_BACKEND=oystehr) are verified synthetic-evaluation adapters: each passed both contract harnesses and the FHIR Agent Safety Eval against a disposable synthetic target. Other FHIR R4 backends need an adapter before they are supported. Epic and Oracle Health sandbox adapters are flag-gated (LASTEHR_EXPERIMENTAL_FHIR_ADAPTERS=1) and can post an approved Observation; they are not a supported path because those sandboxes do not support delete. See docs/adapters/epic-cerner.md and the support matrix for the web, SMART, MCP, auth, and evaluation boundaries before choosing a path.

Self-hosted demos can offer a backend picker (swap the EHR under the live agent, per session) and an "under the hood" panel streaming the agent's FHIR operations — both off by default; see the support matrix for the eligibility rules and npm run check:backends for preflight.

What it isn't

• Not a charting EHR, not a system of record, and not a Medplum replacement. It's a thin agent layer: a small, growing set of tools over your FHIR backend.
• Not a guarantee. The approval gate is a human-in-the-loop boundary, not a safety proof: it stops unilateral writes, but it relies on you reading what you approve. Models can propose wrong or fabricated clinical facts, and approval fatigue is real. Last EHR provides the gate; you provide the review.

How it works

Next.js 15 (App Router) + React 19. The agent lives in app/api/chat/route.ts (streamText + FHIR tools); the FHIR calls go through a small backend interface (lib/fhir/backend.ts) with five built-in adapters: Medplum (hosted or self-hosted, token-authenticated), plus local HAPI FHIR, Firely Server, Aidbox, and Oystehr for synthetic evaluation (docs/support.md). Epic and Oracle Health sandbox adapters are behind LASTEHR_EXPERIMENTAL_FHIR_ADAPTERS and are not part of that supported set. The interface is four methods plus contract notes, so an adapter for another headless EHR is a small, well-scoped contribution. See docs/adapters.md and the roadmap.

flowchart LR
    B["Browser chat"] --> A["/api/chat<br/>streamText + FHIR tools"]
    A -- "reads<br/>search_patients, show_patient_info, read_chart_section" --> M[("Your FHIR backend<br/>(Medplum, HAPI, Firely, Aidbox, Oystehr)")]
    A -- "writes<br/>add_note, record_observation,<br/>record_superseding_observation, create_task" --> C{"Approval card"}
    C -- "Approve & save" --> M
    C -- "Cancel" --> N["Nothing saved"]
    M -. "Authenticated calls use the signed-in user,<br/>bounded by your AccessPolicy.<br/>Chart context reaches the configured model." .-> A

On the public demo, writes are also tagged with your session, so you only ever see the seed data plus your own edits.

Quickstart

The published MCP package and the no-account local HAPI stack are in Start here. The paths in this section are the hosted demo, a Vercel deploy, and a local web app. Longer local-lab notes, including failures, are in docs/quickstart.md.

Hosted synthetic demo (browser, no install): lastehr.com/demo. Synthetic data, no account.

Deploy with Vercel and your own keys:

You'll still need a Medplum project seeded with the synthetic patients (npm run seed, below) for the demo to have data.

Run it locally (authenticated Medplum path). Prerequisites: Node 22.18+ (or 24.2+), a Medplum project (Medplum-hosted free tier or your own), and one tool-capable model API key (OpenAI, Anthropic, or Amazon Bedrock). For a real agent without Medplum, the same .env.local works with FHIR_BACKEND=hapi against the local Docker stack, or with any verified synthetic-evaluation adapter (support matrix).

git clone https://github.com/cbetz/last-ehr.git
cd last-ehr
npm install
cp .env.example .env.local      # then edit .env.local (see below)
npm run seed                     # load synthetic patients into your Medplum
npm run dev                      # http://localhost:3000/demo

At minimum set, in .env.local:

• a model key: OPENAI_API_KEY (default provider), or ANTHROPIC_API_KEY with AI_PROVIDER=anthropic, or AWS credentials plus AI_PROVIDER=bedrock and MODEL_ID;
• NEXT_PUBLIC_MEDPLUM_BASE_URL / MEDPLUM_BASE_URL if you're pointing at your own Medplum (leave blank to use Medplum's hosted API);
• MEDPLUM_CLIENT_ID + MEDPLUM_CLIENT_SECRET (a Medplum ClientApplication): used by npm run seed, and by the no-sign-in quickstart when you also set NEXT_PUBLIC_QUICKSTART=true. Or set NEXT_PUBLIC_MEDPLUM_GOOGLE_CLIENT_ID to sign in via Medplum's Google OAuth instead.

npm run seed loads a small synthetic patient set (scripts/fixtures/patients.ts: four patients with conditions, medications, allergies, immunizations, and vitals/labs, two named "Smith"). It wipes and recreates those patients each run, so it is safe to re-run. Then open /demo and ask: "find patients named Smith." Use synthetic data only.

Local FHIR stack, no Medplum account or model key: the repo ships a Docker Compose stack with HAPI FHIR and Postgres, plus an explicit scripted walkthrough that exercises the approval loop without calling an external model provider.

npm install
npm run demo:local                            # HAPI, seed data, and http://localhost:3000/demo

demo:local forces the local HAPI + scripted configuration without creating or overwriting .env.local. This fixed, no-network walkthrough always finds the seeded Maria Garcia record and proposes one Heart rate: 72 bpm Observation. It does not interpret your prompt, browse charts, or call a model provider; the FHIR wrapper permits only that synthetic record and observation. Press Ctrl-C to stop Next.js; use npm run demo:local:down to remove the local stack. Honest scope: the local HAPI server runs with no auth, so this mode is for local, single-tenant use only; per-browser session isolation is client-side filtering, not a security boundary. The published @lastehr/mcp package targets Medplum (credentialed) or a local HAPI stack; the separate checkout-only npm run mcp:demo Local Lab exposes four synthetic, read-only fixture charts without credentials. Neither local route is a real-data or production path. To run a real agent against HAPI, configure .env.local with a supported model key instead. The hosted public demo stays on Medplum.

To run the app container too, use npm run docker:local after filling .env.local; it combines the HAPI/Postgres compose stack with the app image. Release tags and manual publish runs push a prebuilt scripted-demo image to ghcr.io/cbetz/last-ehr; if the pull fails, no public image has been published yet. See Pull and run from GHCR.

For the longer version, see docs/quickstart.md.

Docs

• Approval-Gated Agent Writes on FHIR: the v0.1 draft protocol — Proposal, Decision, Commit, Audit, and the policy rules.
• Protocol conformance suite: test an implementing MCP stdio server with a scripted reviewer and independent FHIR verification.
• Quickstart: hosted demo, Medplum local run, local HAPI evaluation, and local-lab troubleshooting.
• Support matrix: exactly which backends and interfaces work today.
• Architecture: the chat route, tools, backend adapters, and data boundary.
• Backend adapters: the adapter contract, harnesses, checklist, and contribution path.
• Adapter starter: an executable bearer-token FHIR REST baseline with a contract suite.
• Approval-gated writes: what the gate protects and what it does not.
• MCP server: published chart tools (Medplum, or the local HAPI stack) — read-only by default with an opt-in human-approved write profile — plus the checkout-only synthetic Local Lab.
• FHIR Agent Safety Eval: a disposable synthetic workflow report for proposal, approval, denial, association, and cleanup mechanics.
• Deployment: env vars, rate limiting, Docker, and public-demo hardening.

Launch from the Medplum app (SMART on FHIR)

Last EHR can launch directly from app.medplum.com. Register it once in your Medplum project by creating a ClientApplication with:

• launchUri = https://<your-deploy>/launch
• redirectUri = https://<your-deploy>/launch/callback

then set SMART_CLIENT_ID to that ClientApplication's id in your deployment. Last EHR appears on the Apps tab of every Patient and Encounter page; launching it opens the chat already scoped to that patient, reusing your Medplum sign-in (SMART App Launch with PKCE, public client, no secret). The token is bounded by the granted SMART scopes and your AccessPolicy, and writes still stop at the approval card.

Use chart tools over MCP

The no-account setup for published @lastehr/[email protected] is Start here. This section is the same package with a Medplum credential.

@lastehr/mcp is a small, standalone MCP server for Medplum (or the local HAPI stack). It gives Claude Code, Cursor, and other MCP clients four bounded chart read tools — search_patients, show_patient_info, read_chart_section, and read_document, the same read surface the web agent has; writes exist only as an explicit opt-in where a human approves each proposed write through the client's approval prompt (details). It speaks stdio by default; since 0.4.0 an opt-in HTTP transport lets a hosted agent reach it as an OAuth resource server, with each caller's own FHIR credential obtained per session. The process needs no shared FHIR credential; each session holds its exchanged token (design and evidence).

npx -y @lastehr/mcp init --client claude-code

Use npx -y @lastehr/mcp --version to check the installed MCP package version without starting the server.

Authenticate with a least-privilege MEDPLUM_ACCESS_TOKEN, or MEDPLUM_CLIENT_ID plus MEDPLUM_CLIENT_SECRET; set MEDPLUM_BASE_URL for a self-hosted Medplum instance. Read access can still return PHI, so review the MCP client's model-provider and data-handling boundary before connecting it to a real project.

From a checkout, npm run mcp builds and starts that same package. Full setup, client configuration, and the support boundary are in the MCP guide. The Official MCP Registry listing is the canonical discovery record for the exact, verified release.

For a separate reviewer inbox and work that survives an assistant disconnect, use the checkout-only Work MCP connector. Its four tools read one invented chart, propose a Task, check a durable Work ID, and cancel pending work. It is a local synthetic experiment with ordinary MCP tools; it is not included in @lastehr/mcp and does not implement the MCP Tasks extension or remote OAuth transport.

To evaluate the MCP surface before configuring Medplum, use the separate local lab from a repository checkout:

npm install
npm run mcp:demo -- --client claude-code

It starts local HAPI, reloads the repository's synthetic fixture records, and prints a client registration command. The generated MCP process exposes two of the package's four read-only tool names — search_patients and show_patient_info only, since its fixture client serves too few resource types for the section reader — and can only resolve those fixture patients. It is not included in @lastehr/mcp, does not make HAPI a supported package backend, and must never be pointed at real data. The Local Lab server needs no FHIR credential or provider API key, but the MCP client itself still needs its normal model account and may transmit the synthetic tool output to that model.

Configuration

Every variable is documented in .env.example. The real-agent model path is provider-agnostic with one hard rule: every external provider offered here can carry a BAA, because deployments of this project head toward real clinical data even though the demo is synthetic-only. The scripted option is a separate, explicit, zero-key local HAPI walkthrough, not a model provider or a PHI-ready mode.

AI_PROVIDER	Key	Example MODEL_ID	BAA path
scripted	None	Fixed local synthetic sequence	No external provider call. Requires explicit local HAPI flags; cannot act as a general agent.
openai (default)	OPENAI_API_KEY	gpt-4.1-mini	OpenAI signs BAAs with zero-retention options for API traffic on qualifying plans.
anthropic	ANTHROPIC_API_KEY	claude-sonnet-4-6	Anthropic signs BAAs with zero-retention options for API traffic on qualifying plans.
bedrock	AWS credential env + AWS_REGION	us.anthropic.claude-haiku-4-5-20251001-v1:0	Amazon Bedrock is available under an AWS BAA for eligible services; set an explicit model id or inference profile.

The BAA requirement applies to the external-provider rows: signing one is the operator's step, not a default, and a bare API key is not PHI-ready. Aggregators that cannot sign a BAA (OpenRouter, per their current public terms) are deliberately not offered. Analytics (PostHog) and the marketing-site waitlist (Neon) are optional and lastehr.com-specific.

Security & data

The web agent has no separate chart database, but it processes chart data. The opt-in local Work Lab also stores synthetic requests, run progress, proposals, decisions, and receipts in SQLite. The approval gate is a write-safety control, not a privacy control: in external-model modes, chart context is sent to your configured model provider before write approval. The Work Lab's real agent mode sends your synthetic prompt and bounded invented chart fields to that provider. Its default manual client and test-only scripted mode make no model request. None of these lab modes is qualified for real patient data.

Use synthetic data unless you have real agreements in place. Pointing this at real PHI requires, at minimum:

• a BAA with your model provider that covers your API traffic (consumer plans don't qualify);
• a HIPAA-eligible FHIR backend with its own BAA;
• your own compliance review. This project is alpha and is not a HIPAA-covered service; PHI handling is the operator's responsibility.

See SECURITY.md for the full posture and how to report vulnerabilities.

Open-core

Self-hosting is free and Apache-2.0. A managed hosted tier (managed Medplum + a signed BAA, multi-tenancy, billing) may follow, built only after the open-source core has traction.

Not affiliated

A personal open-source project. Not affiliated with, endorsed by, or sponsored by Medplum, Vercel, or any employer.

License

Apache-2.0. See NOTICE for third-party attributions, CONTRIBUTING.md to contribute, and GOVERNANCE.md for how the project is run.