meterbility-agent 0.3.1

Meterbility SDK for Python — trace LLM agent runs into the local Meterbility store

meterbility-agent (Python SDK)

The Python counterpart to @meterbility/agent. Wire any Python LLM agent into Meterbility by capturing one Step per model call. Runs land in the same ~/.meterbility/meterbility.db store the TypeScript SDK, the CLI, and the web UI read from — so a Python agent shows up in meter list and meter web immediately, with no separate ingest step.

Install

pip install -e packages/agent-py        # core only (stdlib)
pip install -e 'packages/agent-py[anthropic]'  # add the Anthropic helper

Requires Python 3.9+. Core has zero runtime dependencies (it uses sqlite3, hashlib, json, pathlib from the standard library).

Minimal usage

from meterbility_agent import MeterbilityTracer

with MeterbilityTracer(project="my-app", agent="support") as tracer:
    step = tracer.start_step(
        model="claude-opus-4-7",
        system_prompt="you are helpful",
        history=[{"role": "user", "content": "hello"}],
    )
    # ... call your model, then record what it did ...
    step.record_message("hi back!").record_tokens(
        input=10, output=2, cached_read=0, cache_creation=0
    )
    step.end()

Then in another terminal:

meter list
meter inspect <run-id>
meter web   # see it live alongside any Claude Code / Codex runs

Anthropic shortcut

If you already use the official anthropic Python SDK, wrap the client once and every messages.create() call is captured automatically:

from anthropic import Anthropic
from meterbility_agent import MeterbilityTracer, trace_anthropic

with MeterbilityTracer(project="my-app", agent="support") as tracer:
    client = trace_anthropic(tracer, Anthropic())  # API-compatible wrapper

    resp = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=512,
        system="you are helpful",
        messages=[{"role": "user", "content": "hello"}],
    )
    # one Meterbility Step captured per call — including tool_use, tokens, latency.

The wrapped client behaves identically to the underlying Anthropic client — it returns the same Message object, exposes the same attributes, and re-raises the same exceptions (recording them as error outcomes before the throw bubbles).

Multi-step agent loop

For multi-step agents that call tools, drive each step manually:

from meterbility_agent import MeterbilityTracer, tool_call_action

with MeterbilityTracer(project="my-app", agent="rag") as tracer:
    history = [{"role": "user", "content": "what's the weather in NYC?"}]

    # turn 1: model picks a tool
    step1 = tracer.start_step(model="claude-opus-4-7", history=history)
    step1.record_action(tool_call_action("weather", {"city": "NYC"}, "tu_1"))
    step1.record_tokens(input=120, output=20)
    step1.end()

    # run the tool yourself ...
    tool_result = {"temp_f": 68, "conditions": "clear"}

    # turn 2: feed the tool result back, model produces a message
    history += [
        {"role": "assistant", "content": "[tool_call weather]"},
        {"role": "tool", "content": str(tool_result)},
    ]
    step2 = tracer.start_step(model="claude-opus-4-7", history=history)
    step2.record_message("It's 68°F and clear in NYC.")
    step2.record_tokens(input=150, output=15)
    step2.end()

What's captured per step

Field	Source
model	required arg to start_step()
system_prompt, tool_definitions, history, retrieved_docs	optional context components (blob-stored, content-addressed)
action	record_message() / record_tool_call() / record_action()
outcome	record_tool_result() / record_outcome()
tokens (input / output / cached_read / cache_creation [5m] / cache_creation_1h)	record_tokens()
cost_cents	computed from tokens via the same pricing table the TS side uses
latency_ms	wall-clock since start_step(), or explicit via record_tokens(latency_ms=...)
tags	step.tag("x") — cost:approx auto-applied for unknown models

Configuration

Env var	Effect
METERBILITY_HOME	Override the data directory (default ~/.meterbility)
METERBILITY_REDACT=off	Disable regex redaction of API keys / bearer tokens / private keys

Run the tests

cd packages/agent-py
python -m unittest discover -s tests -v

Tests redirect METERBILITY_HOME to a tempdir per test, so they never touch your real ~/.meterbility store.