agentwatch-observe

agentwatch-observe

Lightweight observability for AI agents. Wrap any local or cloud model call, stream steps to AgentWatch, and get cost tracking, anomaly detection, and guardrails on the dashboard at agentwatch.in.

Privacy: Your model credentials (OpenAI, Gemini, Ollama, etc.) stay on your machine. AgentWatch only receives telemetry — tokens, cost estimates, tool names, timing, and step metadata — never your API keys or raw model weights.

Install

pip install agentwatch-observe

Optional LangChain integration:

pip install "agentwatch-observe[langchain]"

Quickstart (any agent, any model)

Add to your project .env:

AGENTWATCH_API_KEY=your_key_here
AGENTWATCH_BASE_URL=https://api.agentwatch.in   # optional

Two lines in your agent code — works with Gemini, OpenAI, Ollama, LangChain, or custom logic:

import agentwatch

agentwatch.bootstrap()  # reads AGENTWATCH_API_KEY from env

@agentwatch.track(agent_name="my-agent")
def handle_message(message: str):
    return your_existing_agent_logic(message)

Open the AgentWatch dashboard → Runs to see the trace.

Optional: auto-capture provider SDK calls

If you call a provider SDK directly, wrap it once and every LLM call is logged:

from google import genai
client = agentwatch.instrument_google_genai_client(genai.Client(api_key="..."))
# client.models.generate_content(...) is now auto-logged inside an active run

# Also: instrument_openai(OpenAI()), instrument_anthropic(Anthropic())

Manual step logging (advanced)

with agentwatch.run("my-agent") as run:
    run.log_llm(model="gpt-4o-mini", input_tokens=120, output_tokens=40, output="Hello!")
    run.log_tool("search", input_data={"q": "hello"}, output_data={"results": []})

Framework-agnostic API

• agentwatch.run(...) — context manager
• run.log_llm(...) — LLM step with token/cost tracking
• run.log_tool(...) — tool call step
• agentwatch.track(...) — decorator
• agentwatch.instrument_openai(client) — auto-log OpenAI calls

See the full docs at agentwatch.in.

Releasing a new version

PyPI never allows re-uploading the same version number.

1. Bump version in pyproject.toml and agentwatch/__init__.py
2. cd sdk && python -m build
3. twine check dist/*
4. Upload to TestPyPI first: twine upload --repository testpypi dist/*
5. After verification, upload to PyPI: twine upload dist/*

See CONTRIBUTING.md for details.