Appearance

Quickstart ​

AgentSight records what your agent did — the exchanges, the answer latency, the tool calls, the LLM spend — from the smallest amount of code that can honestly capture it. No infrastructure, no collector to run.

Installation ​

bash
pip install agentsight python-dotenv

python-dotenv is included because with it present, importing agentsight picks up a .env file — so your key never has to be in code.

Or let your coding agent do it ​

If a coding agent works in your repository, there is a packaged skill that makes the whole integration a task it completes correctly on the first pass — it reads your code first, asks only what code cannot answer, and verifies what it recorded before claiming success. Install it:

bash
npx skills add agentsightio/agentsightio

then ask in plain words — "add AgentSight to this service" — and skip the rest of this page. What the skill does, the agents it works with, and the no-installer route are on Use with AI agents.

Setup ​

Get an API key from the AgentSight dashboard and put it in .env:

bash
AGENTSIGHT_API_KEY="your_api_key_here"

Message content is transmitted

Everything you track — message text, tool arguments, tool responses — is sent to AgentSight and stored. Nothing is scraped or inferred, but you should know exactly what leaves your process before you ship. What the SDK sends lists every field, and shows you how to dump the payload locally without sending it.

Quickstart ​

python
import agentsight

agentsight.init()

@agentsight.turn(id_from="session_id", infer=True)
def handle_message(session_id: str, text: str) -> str:
    return my_agent.run(text)

That's a complete integration. Every call records a conversation, both messages and the answer latency — plus every LLM call made inside it, with token usage and cost, and every tool call the SDK can see: the functions you decorate with @agentsight.tool and the tools your framework reports.

When this shortcut applies

infer=True reads the user message from the first string argument and the agent message from the return value. It fits handlers shaped like (text) -> str.

Most production handlers aren't — they take a framework request object, or the id is nested in a payload, or the text is rewritten before the agent sees it. For those, see Turns & Messages, which is a few more lines and works everywhere.

id_from names the parameter holding your conversation id. When the id is nested somewhere less convenient, it also takes a callable:

python
@agentsight.turn(id_from=lambda args: json.loads(args["data"])["conversation_id"])
def chat(request: Request, data: str = Form(...)):
    ...

Seeing it land ​

Your conversations appear in the dashboard within a few seconds.

If you would rather look before anything is transmitted, point the SDK at a directory instead of the network — no key, no account, no requests:

bash
AGENTSIGHT_FILE_EXPORTER=./agentsight-traces python your_app.py

AGENTSIGHT_FILE_EXPORTER names a directory, not a file — the SDK creates it if needed and writes one JSON file into it per batch. (Putting the variable in front of python your_app.py sets it for that one run; exporting it or adding it to .env works the same.) You get the exact JSON the API would have received.

Next steps ​