clarismd

Official Python SDK for the ClarisMD governed-AI gateway for healthcare. One-line drop-in replacement for the OpenAI SDK with PHI detection, audit logging, and configurable policy enforcement built in.

Note. ClarisMD is a governance gateway. It does not, by itself, make any call HIPAA compliant — compliance is a property of your full deployment, contracts (BAAs), and operational controls. See NOTICE for the full disclaimer.

Install

From PyPI:

pip install clarismd

From the public GitHub repo (latest main, or pin to a tag/commit):

pip install git+https://github.com/amanjaiswal777/clarismd-python.git
pip install git+https://github.com/amanjaiswal777/clarismd-python.git@v0.1.0
pip install git+https://github.com/amanjaiswal777/clarismd-python.git@<commit-sha>

In a requirements.txt:

clarismd @ git+https://github.com/amanjaiswal777/clarismd-python.git@v0.1.0

For local development (editable install with dev extras):

git clone https://github.com/amanjaiswal777/clarismd-python.git
cd clarismd-python
pip install -e ".[dev]"

Python 3.9+ supported.

Quickstart

from clarismd import ClarisMD

client = ClarisMD()  # reads CLARISMD_API_KEY

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize hypertension treatment options."}],
)
print(resp.choices[0].message.content)

Set your API key once:

export CLARISMD_API_KEY=cmd-...

Migrating from OpenAI

The surface mirrors openai>=1.0. In most cases, the diff is one line:

-from openai import OpenAI
-client = OpenAI()
+from clarismd import ClarisMD
+client = ClarisMD()

Everything else — client.chat.completions.create, client.embeddings.create, streaming, async — works the same way.

Self-host

Point the SDK at your own gateway with one environment variable:

export CLARISMD_BASE_URL=https://gateway.your-org.internal/v1
export CLARISMD_API_KEY=cmd-...

Or pass it explicitly:

client = ClarisMD(base_url="https://gateway.your-org.internal/v1")

Verify the gateway is reachable (no API key required):

status = client.health.check()
print(status.status, status.version)  # "ok" 0.4.2

Streaming

stream = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Explain ACE inhibitors briefly."}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta
    if delta.content:
        print(delta.content, end="", flush=True)

Streams expose .close() and the underlying httpx.Response via stream.response, and can be used as context managers.

Async

import asyncio
from clarismd import AsyncClarisMD

async def main() -> None:
    async with AsyncClarisMD() as client:
        resp = await client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{"role": "user", "content": "Hi"}],
        )
        print(resp.choices[0].message.content)

asyncio.run(main())

PHI scan

Run text through the PHI detector without making a generation call:

result = client.phi.scan("Patient John Doe, MRN 4471, called the clinic.")
if result.detected:
    for entity in result.entities:
        print(entity.type, "->", entity.text)

Audit export

Pull a signed evidence bundle for a date range:

from datetime import datetime, timezone

bundle = client.audit.export(
    format="pdf",
    start_date=datetime(2026, 5, 1, tzinfo=timezone.utc),
    end_date=datetime(2026, 5, 31, tzinfo=timezone.utc),
)
with open("audit-may.pdf", "wb") as f:
    f.write(bundle)

Compliance score

score = client.compliance.score(framework="hipaa")
print(f"Auto-verified: {score.auto_verified.satisfied}/{score.auto_verified.total}")
print(f"Manual:        {score.manual.acknowledged}/{score.manual.total}")

Error handling

All errors derive from ClarisMDError. The closed-set typed errors let you branch precisely:

from clarismd import (
    ClarisMD,
    AuthenticationError,
    PHIPolicyViolationError,
    RateLimitError,
)

client = ClarisMD()

try:
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": "..."}],
    )
except PHIPolicyViolationError as exc:
    # Request was blocked by a configured PHI policy.
    print("PHI violation:", exc, "request_id:", exc.request_id)
except RateLimitError as exc:
    # Retry-After honored automatically; this fires only after the budget is spent.
    print("Backed off too long:", exc)
except AuthenticationError:
    # 401 — bad / missing / revoked key.
    raise

Every APIError carries .status_code, .request_id, .code, .param, .type, and .body for structured handling and bug reports.

ClarisMD-specific request options

Two extra kwargs are accepted on every method and translate into HTTP headers the gateway understands:

client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "..."}],
    clarismd_phi_action="redact",
)

Examples

Runnable scripts live in examples/:

• quickstart.py — five-line basic call
• streaming.py — server-sent-events streaming
• async_chat.py — AsyncClarisMD round-trip
• phi_scan.py — PHI entity detection
• audit_export.py — pull a PDF evidence bundle
• compliance_score.py — auto vs. manual breakdown

License

Apache-2.0. See LICENSE and NOTICE.

The ClarisMD backend is licensed AGPL-3.0; this SDK is intentionally permissive so you can ship it inside any application without copyleft obligations.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md for the dev setup and PR conventions, and CODE_OF_CONDUCT.md for community expectations.

Security

Found a vulnerability? Please report it privately per SECURITY.md — do not open a public issue.