AgenticMeter

Find the bugs that waste money in your agent runs.

AgenticMeter is a linter for AI agent runs. Wrap your agent, run it, and get told what went wrong and what it cost: tool loops, duplicate LLM calls, context bloat, failures your agent quietly swallowed. You don't have to dig through a trace to find them.

It runs locally. Traces stay in a SQLite file on your machine, there is no account or cloud service, and the core has zero dependencies.

Try it in 10 seconds

pip install agenticmeter-cli
agenticmeter demo

demo runs a simulated travel agent (no API keys, no network) with the bugs real agents ship with:

▸ travel_agent — 11 steps · 28.2k tok · $0.0739 · 0.7s

⚠ One step = 69% of this run's cost
    'openai.chat' cost $0.0510 of $0.0739 total.
    → optimize or cache that single call before anything else.

⚠ Tool failed 1x but the run succeeded anyway
    fetch_hotel_prices errored mid-run; the agent recovered, so nothing surfaced.
    → check whether the recovery path is actually correct, not just quiet.

⚠ Context grew 1,200 → 18,000 tokens
    prompt size grew 15.0x across 5 calls (~$0.0420 of re-sent history on that call)
    → trim or summarize history between steps to stop the quadratic creep.

⚠ Same tool called 4x with identical args
    'search_flights' ran 4 times with the same input — likely a loop or a missing cache.
    → add a cache, or fix the condition that re-calls it.
...

Every warning links to the exact spans that prove it. Nothing is guessed.

Lint your own agent

import agenticmeter as am

am.configure(sink="sqlite")   # persist runs to ~/.agenticmeter/traces.db

@am.tool
def search(query): ...

@am.meter
def agent(task):
    ...

agent("plan my launch week")  # warnings print to stderr when the run ends

OpenAI, Anthropic and LangChain calls are captured automatically when those libraries are installed, with tokens and cost per call. For anything else, mark steps by hand:

with am.span("retrieve", am.SpanType.RETRIEVAL):
    ...

Clean runs stay silent. Pass summary=True to always print, or summary=False to never print.

What it catches

Thresholds live in agenticmeter.analysis.insights.Config. Very short runs (fewer than two steps, or under 50ms of work) don't trigger the time check.

Look back at runs

agenticmeter runs            # recent runs: steps, tokens, cost
agenticmeter show            # latest run: trace tree + warnings
agenticmeter show <id>       # a specific run (id prefix is fine)
agenticmeter ui              # local web viewer at http://127.0.0.1:4319
agenticmeter demo --save     # keep the demo run so you can explore it in `show` / `ui`

In the viewer, clicking a warning highlights the steps that caused it.

Design rules

• Zero-dependency core. Integrations are optional extras: pip install "agenticmeter-cli[openai,anthropic,langchain]".
• Local-first. No telemetry, no network calls. Your traces never leave your machine.
• Fail open. If AgenticMeter hits an internal error, it drops the span and logs once. It never crashes or slows your agent.
• Secrets scrubbed by default. Inputs and outputs are redacted before they are stored (redact="scrub"; also "truncate", "off" or your own function).
• Prices are configurable. Override or add models with am.cost.set_prices({...}).

Install notes

The PyPI package is agenticmeter-cli. The import name and command are both agenticmeter. Requires Python 3.10+.

Development

pip install -e ".[openai,anthropic]"
python smoke_test.py       # capture layer
python behavior_test.py    # detectors: messy run fires all, clean and trivial runs fire none

See CONTRIBUTING.md. MIT licensed.