To explore a synthetic health record before installing, open the
Mirobody Web demo.
A · Use the library
Install uv, then run:
血红蛋白 and ヘモグロビン both resolve to LOINC 718-7. The broad category
血脂 is left unresolved to avoid assigning a code for the wrong test.
For Python 3.12+ projects, install the library:
pip install 'mirobody[parse]'
and configure a model API key. For an MCP client, run
uvx --python 3.12 mirobody mcp: vocabulary tools work offline over stdio;
the document tool standardize_report requires [parse] and a model key.
See standardization.md for symptom and diagnosis coding,
unit conversion and FHIR output, or try the scripts in
examples/.
B · Start the full app
The app includes the web client, server, background worker and Postgres with pgvector. Choose a cloud or local setup.Try it in Codespaces
18060 is forwarded privately by default.
The app and records are stored in GitHub’s cloud; documents and questions sent
for model processing go to your model provider. Use the supplied synthetic
data to explore. Codespaces billing
and model charges apply. Stop the codespace when finished; reopening it starts
the app again.
Run it 100% locally
With Git, OpenSSL and Docker Compose installed:.env, selects the local models and prints
the app link. The prebuilt image includes the LOINC bundle, so Git LFS is not
needed for Docker deployment.
To use a cloud model, run ./deploy.sh without a local profile. Open the
setup link at http://localhost:18060 and enter a model API key. Records stay
on your host; documents and questions sent for model processing go to that
provider.
Sign in and try an upload
On the Email code tab, sign in asyou@mirobody.ai with code 111111.
No mail service is needed. Ask “How has my cholesterol changed?”, then upload
a synthetic file from demo/upload/ on the Data page. Those files
are not included in the seeded records, so uploading one tests document processing.
Before the first start of a deployment for real data, set
SEED_DEMO_DATA=false. Otherwise the app creates the two public demo accounts.
You can also put a model key in .env: OPENROUTER_API_KEY, OPENAI_API_KEY,
GOOGLE_API_KEY or ANTHROPIC_API_KEY. Use <PROVIDER>_BASE_URL for an
OpenAI-compatible gateway. config.llm.yaml lists the
variable each entry reads; store secrets in the environment, not that file.
After changing .env, run docker compose up -d; restart does not reload it.
Check the selected models and send a test request to each:
C · Develop from source
version. Run dev from the repository root to serve
the web client in frontend/ alongside the API. The Python package alone does
not include the web client.
dev generates temporary secrets. To preserve sessions and encrypted
configuration across restarts, set JWT_KEY and CONFIG_ENCRYPTION_KEY. Set
PG_ENCRYPTION_KEY as well: without it, stored record fields are encrypted
under config.yaml’s public placeholder (SECURITY.md).
For deployment, serve reads the shipped configuration and
config.{ENV}.yaml and expects persistent secrets.
On Debian or Ubuntu, scripts/install_native.sh installs PostgreSQL, pgvector
and a Python environment without Docker; scripts/start_native.sh starts the
worker and server on port 18060.