Supported versions
SDK 1.0.0: the runtimes it needs, what each provider and framework gives you, and the library versions we've verified it against.
Both SDKs are at 1.0.0, published as token-police.
| Python | Node | |
|---|---|---|
| Runtime | Python 3.10+ | Node 20+ (ESM and CommonJS) |
| Install | pip install token-police | npm i token-police |
| Import | import token_police as tp | import * as tp from "token-police" |
The SDKs deliberately pin no provider library version. They detect the shape of whatever
you have installed at runtime and fail open if it's something they don't recognise — so a
provider release can't break your app, it can only cost you metering until we catch up. The
one place that shows through today is Node + openai 7 (below).
Python
Everything in this table works on the base pip install token-police, except where an extra
is named.
| Provider / framework | Metered | Streaming | Tool spans | Install |
|---|---|---|---|---|
| OpenAI | auto | ✅ | via framework | base |
| Anthropic | auto | ✅ | via framework | base |
| Amazon Bedrock | auto | ✅ | via framework | base |
| Google Gemini | auto | ✅ | via framework | base |
| Cohere | ✅ | ✅ | — | base (chat and embeddings; no extra exists) |
| Mistral | ✅ | ✅ | — | base (chat, embeddings, OCR, transcription, FIM, speech on 2.x) |
| Together · Groq · Cerebras · OpenRouter · Hugging Face · LiteLLM | ✅ | ✅ | — | base |
| xAI | ✅ | ✅ | — | base — [xai] installs the xai-sdk client itself |
| Voyage AI | ✅ | n/a | — | [voyageai] (embeddings) |
| LangChain | ✅ | ✅ | auto | [langchain] |
| LlamaIndex | ✅ | ✅ | auto | [llamaindex] |
| Pydantic AI | ✅ | ✅ | auto | [pydantic-ai] |
| OpenAI Agents | ✅ | ✅ | auto | [openai-agents] |
| CrewAI | ✅ | ✅ | manual | [crewai] |
| Agno | ✅ | ✅ | manual | [agno] |
Extras: all, langchain, llamaindex, pydantic-ai, openai-agents, crewai, agno,
xai, voyageai. (google also resolves, but it's an empty no-op kept for back-compat —
Gemini is metered by the base install.)
pip install "token-police[all]" bundles the framework extras except CrewAI. CrewAI
supports Python 3.10–3.13 but not 3.14 yet, so keeping it out of all is what lets [all]
install everywhere; add it on its own line: pip install "token-police[crewai]".
Verified with
Python apps in our test fleet pin floors rather than exact versions, so most of these are "at least this" rather than "exactly this".
| Library | Verified |
|---|---|
openai | floor ≥ 1.66.0 |
anthropic | ≥ 0.111.0, < 1 |
google-genai | 2.14.0 |
boto3 (Bedrock) | floor ≥ 1.39.0 |
cohere | floor ≥ 5.13.0 (5, 6 and 7 all work) |
mistralai | 1.12.4 and 2.9.4 (both majors supported — see the note below) |
langchain-core / -openai / -anthropic / -google-genai | 1.5.2 / 1.4.1 / 1.4.6 / 4.2.5 |
langgraph | 1.2.10 |
llama-index-core / -llms-openai / -llms-google-genai | 0.14.23 / 0.7.10 / 0.9.6 |
together / litellm / pydantic-ai / openai-agents / agno / voyageai / huggingface_hub / cerebras_cloud_sdk / xai-sdk | floors only (≥ 1.3.0 / ≥ 1.50.0 / ≥ 0.5.0 / ≥ 0.1.0 / ≥ 2.0.0 / ≥ 0.2.0 / ≥ 0.30.0 / ≥ 1.0.0 / ≥ 0.1.0) |
- Mistral: no pin needed —
mistralai1.x and 2.x both work. Version 2 moved the client import tofrom mistralai.client import Mistral(1.x keepsfrom mistralai import Mistral). Onmistralai2.0–2.9.1, pip may install an older OpenTelemetry train to satisfy that release's own cap; it still meters normally, andmistralai>=2.9.2dropped the cap. - Never set
OTEL_SDK_DISABLED=true— it turns OpenTelemetry off globally, TokenPolice's own capture included.
Node
Everything is in the one npm i token-police install. Install the provider package you
already use (they're optional peer dependencies) and TokenPolice taps it.
| Provider / framework | Metered | Streaming | Tool spans | Notes |
|---|---|---|---|---|
OpenAI (openai) | auto | ✅ | via framework | On openai 7.x, pass the module — see below |
Anthropic (@anthropic-ai/sdk) | auto | ✅ | via framework | Prefer the namespace import in instrumentModules |
Cohere (cohere-ai) | ✅ | ✅ | — | v2 client |
Amazon Bedrock (@aws-sdk/client-bedrock-runtime) | ✅ | ✅ | — | Converse usage read directly |
Google Gemini (@google/genai) | ✅ | ✅ | — | 1.x and 2.x |
| OpenRouter · Cerebras · Together · Groq · Mistral | ✅ | ✅ | — | each provider's own package |
Hugging Face (@huggingface/inference) | ✅ | ✅ | — | must be wired before you construct InferenceClient |
Vercel AI SDK (ai + @ai-sdk/*) | ✅ | ✅ | manual — always | community providers need instrumentModules.aiSdkProviders |
| LangChain | ✅ | ✅ | auto | add the token-police-langchain companion |
LlamaIndex (@llamaindex/*) | ✅ | ✅ | auto | ESM: pass each provider namespace under instrumentModules.llamaIndex |
OpenAI Agents JS (@openai/agents) | ✅ | ✅ | auto | runs on the Responses API |
openai 7.x — pass the module. TokenPolice's built-in OpenAI capture auto-discovers
openai 4–6. On 7.x, auto-discovery declines the package and prints a warning at startup:
chat-completion calls still run and are still pre-flight checked, but they report 0 tokens,
so budgets never accumulate. Passing the module fixes it completely:
import OpenAI from "openai";
tp.init({ /* … */ instrumentModules: { openAI: OpenAI } });That patches your own copy directly, and chat completions meter normally on openai 7 with
the right model, token counts, and composition. The Responses API, images, audio, and
embeddings are metered on openai 7 either way. See Node & ESM.
Verified with
These come from the lockfiles of the apps we run the SDK against, so they're exact resolved versions.
| Library | Verified |
|---|---|
openai | 7.9.0 (with instrumentModules), 4.104.0 |
@anthropic-ai/sdk | 0.123.0 |
@google/genai | 2.20.0 |
@aws-sdk/client-bedrock-runtime | 3.1124.0 |
ai (Vercel AI SDK) | 7.0.90 |
@ai-sdk/anthropic / -gateway / -xai / -openai-compatible | 3.0.6 / 4.0.72 / 2.0.92 / 1.0.53 |
@langchain/core / -openai / -anthropic / -google-genai / langgraph | 1.2.9 / 1.5.11 / 1.5.9 / 2.3.0 / 1.4.13 |
llamaindex / @llamaindex/openai / -anthropic / -google | 0.12.1 / 0.4.22 / 0.3.26 / 0.4.0 |
cohere-ai / @mistralai/mistralai / together-ai | 8.1.0 / 2.6.4 / 0.50.0 |
@openai/agents / @huggingface/inference | 0.17.0 / 4.13.28 |
@cerebras/cerebras_cloud_sdk / @openrouter/sdk | 1.91.0 / 1.2.93 |
groq-sdk is supported by the code but isn't covered by a running app in our fleet, so treat
it as untested at that tier.
Beyond chat
Embeddings, image generation, text-to-speech, transcription, video, and OCR are metered too,
each tagged with its own operation so a rule can target it separately from chat.
Known gaps — these are not metered today: rerank endpoints (Cohere, Voyage), moderation
endpoints, and OpenAI's images.edit / images.createVariation. If you need them counted,
wrap the call yourself with protect() and a provider hint. Never
protect() something already in the tables above — you'll count it twice.