tolvyn

Drop-in replacement for openai and anthropic. One line change. Every AI call metered, attributed, and governed.

10,000 free requests forever. No credit card.

Install

pip install tolvyn

Python 3.9 or later required.

For Google support:

pip install "tolvyn[google]"

Quick start

# Before
from openai import OpenAI
client = OpenAI()

# After — one line change
from tolvyn import OpenAI
client = OpenAI(
    tolvyn_api_key="tlv_live_...",
    team="backend",
    service="summariser",
)

# Everything else stays the same
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

All three providers

from tolvyn import OpenAI, Anthropic, Google

# OpenAI
oai = OpenAI(
    tolvyn_api_key="tlv_live_...",
    openai_api_key="sk-...",            # optional — enables fail-open fallback
)

# Anthropic
anth = Anthropic(
    tolvyn_api_key="tlv_live_...",
    anthropic_api_key="sk-ant-...",     # optional — enables fail-open fallback
)

# Google (requires the [google] extra)
goog = Google(tolvyn_api_key="tlv_live_...")
model = goog.GenerativeModel("gemini-1.5-flash")

Async

AsyncOpenAI and AsyncAnthropic mirror the sync classes:

import asyncio
from tolvyn import AsyncOpenAI, AsyncAnthropic

async def main():
    openai_client = AsyncOpenAI(tolvyn_api_key="tlv_live_...")
    anth_client = AsyncAnthropic(tolvyn_api_key="tlv_live_...")

    openai_resp, anth_resp = await asyncio.gather(
        openai_client.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": "Hello"}],
        ),
        anth_client.messages.create(
            model="claude-haiku-4-5",
            max_tokens=1024,
            messages=[{"role": "user", "content": "Hello"}],
        ),
    )

asyncio.run(main())

There is no AsyncGoogle class — Google's SDK uses process-global configuration, so async calls go through the same Google instance via generate_content_async.

Attribution headers

Set any combination of these on construction; the SDK sends them as X-Tolvyn-* headers automatically:

client = OpenAI(
    tolvyn_api_key="tlv_live_...",
    team="backend",
    service="invoice-summarizer",
    feature="summarize",
    agent="claude-code",
    user="alice@company.com",
    end_customer="acme-corp",
)

The TOLVYN proxy strips all six headers before forwarding the request upstream — they never reach OpenAI/Anthropic/Google.

Fail-open behavior

If TOLVYN's proxy is unreachable, the SDK automatically retries the request directly against the provider (requires openai_api_key / anthropic_api_key to be set). Disable with fail_open=False.

Triggers on: httpx.ConnectError, ConnectTimeout, ReadTimeout, WriteTimeout, RemoteProtocolError, and HTTP 503 from the proxy. Does NOT trigger on: 4xx errors (auth failures, rate limits, bad requests).

Requests that fail open bypass the proxy and are not metered for that call.

Google limitations

The Google class accepts fail_open for API parity but fail-open is not yet supported for Google due to an upstream google-generativeai SDK limitation (no httpx transport hook). The SDK emits a UserWarning when fail_open=True is passed to Google so callers know it won't trigger. Pass fail_open=False to silence.

google-generativeai also uses process-global configuration via genai.configure(). Creating a second Google instance in the same process overwrites the first's config — the SDK emits a UserWarning in that case. Use one Google per process.

Environment variables

export TOLVYN_API_KEY="tlv_live_..."
export OPENAI_API_KEY="sk-..."

from tolvyn import OpenAI
client = OpenAI()   # picks up TOLVYN_API_KEY automatically

API keys

• Production keys start with tlv_live_
• Test keys start with tlv_test_ (use these in CI / staging)
• Get your key at app.tolvyn.io → API Keys
• Provider keys (OpenAI / Anthropic / Google) go in the dashboard under Account → Provider Keys — never in code. They are stored encrypted server-side.

Changelog

github.com/tolvyn/tolvyn-cli/releases

Links

• Docs: docs.tolvyn.io/sdks/python
• Quickstart: docs.tolvyn.io/getting-started/quickstart
• Dashboard: app.tolvyn.io
• Issues: github.com/tolvyn/tolvyn-python/issues

Feedback

founder@tolvyn.io — we read every message.

© 2026 TOLVYN. MIT licensed.