Appearance

Tokens & Cost ​

This is the shortest page in the documentation, because there is almost nothing to do. Every LLM call made inside a conversation scope is counted automatically — token usage, the model that actually answered, and the cost — with no decorator, no wrapper and no counter to keep.

python
with agentsight.conversation("wa-3859"):
    with agentsight.turn():
        agentsight.user_message(text)
        reply = client.chat.completions.create(...)   # counted
        agentsight.agent_message(reply.choices[0].message.content)

What is covered ​

IntegrationHow it attachesWhat it sees
OpenAIpatches the provider clientevery call, including ones a framework makes for you
Anthropicpatches the provider clientevery call, including ones a framework makes for you
LangChainregisters with the callback systemtool calls, plus LLM calls no provider patch covers
LlamaIndexregisters with the dispatchertool calls, plus LLM calls no provider patch covers

Nothing needs configuring. init() installs whichever of the four are importable and skips the rest.

Models reached through a framework rather than through a patched provider client — Bedrock, Vertex AI, Ollama, LiteLLM and the rest — are recorded by the framework integration, which is the only thing watching them. So a LangChain app on Bedrock is counted; the same Bedrock call made directly, with no framework involved, is not.

Running a provider integration and a framework integration together is the normal case and does not double-count. A LangChain app on OpenAI produces one LLM record from the provider side and tool records from the framework side, because the framework side stands down on any call the provider side already sees.

What "every call" means ​

Every call. Including the ones nobody thinks to track: the evaluator scoring the answer, the router picking a path, the classifier deciding whether the question is in scope, the summariser compressing history. Those calls are usually small and always numerous, and they are where surprising cost hides.

That is the argument for scoping by conversation rather than counting by hand. Anything that runs inside the scope is on the books whether or not you remembered it existed.

Prompts and completions are never sent

LLM records carry the accounting — model, token counts, timing — and nothing else. The text of what you sent the model and what it sent back does not leave your process. Message content does, but only the text you pass to user_message() and agent_message() yourself.

Cost ​

Cost is priced on our side, from your exact token counts. The SDK does not carry a rate table and does not compute a price.

That matters more than it sounds. Rates change, and published rates are sometimes wrong; pricing server-side means your history can be restated when either happens, instead of staying frozen at whatever the SDK happened to know on the day you deployed it.

A model we have no rate for is marked unpriced — not zero. Unknown spend reads as unknown, and it can be repriced retroactively once a rate exists for it.

When a model cannot be named ​

Detection reads the model from the provider's own response, or from the arguments of the call. A few wrappers give it neither: a hand-rolled chat model around a self-hosted server, a custom framework LLM that reports nothing about itself.

A call with no model name is permanently unpriceable — no rate can ever match it, and repricing cannot recover what was never named. model_hint() is where you say what is on the other end:

python
with agentsight.model_hint("llama3.1:8b"):
    chain.invoke(...)

It is strictly a fallback: a call that resolves a model of its own keeps it, and the hint fills only the hole it left. Other providers covers it in full, alongside the providers that need it.

Turning it off ​

init() installs everything importable by default. Pass False for none of it:

python
agentsight.init(auto_instrument=False)

Or name the ones you want:

python
agentsight.init(auto_instrument=["openai", "langchain"])

Valid names are openai, anthropic, langchain and llama_index. The spellings llamaindex, llama-index, claude and openai_agents are accepted as aliases.

A misspelled target is silent

An unrecognised name is skipped rather than reported, so auto_instrument=["open_ai"] disables OpenAI capture and says nothing about it. Check the spelling against the list above if usage stops appearing.

Turning an integration off does not turn tracking off — conversations, turns, messages and decorated tools carry on exactly as before. You lose the automatic LLM accounting for that provider, and nothing else.

Next ​