agent_metering

Language-agnostic LLM cost metering for B2B SaaS — point any OpenAI / Anthropic client at the HTTP proxy (Node, Go, Java, PHP, Python, curl, …).

MIT open source — CONTRIBUTING · SECURITY · Code of Conduct

Install

pip install llm-agent-metering
# optional extras:
# pip install "llm-agent-metering[dashboard]"
# pip install "llm-agent-metering[example]"

From GitHub (latest main):

pip install "git+https://github.com/prantakhandaker/agent_metering.git"

Any language (recommended)

Run the proxy once, then set your SDK base URL to it. No SDK install in the app language required.

pip install llm-agent-metering
python -m uvicorn agent_metering.proxy:app --host 0.0.0.0 --port 8787

Optional per-user / feature headers (stripped before upstream):

• X-User-Id (preferred) or X-Customer-Id
• X-Feature
• Or OpenAI body field user / Anthropic metadata.user_id

Defaults: env AGENT_METERING_CUSTOMER_ID / AGENT_METERING_FEATURE, else default.

Spend → local SQLite agent_metering.db. Dashboard:

python -m streamlit run examples/dashboard.py   # pip install "llm-agent-metering[dashboard]"

Node

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "http://127.0.0.1:8787/proxy/openai/v1",
  defaultHeaders: { "X-User-Id": "user_42" },
});

await client.chat.completions.create({
  model: "gpt-4o-mini",
  messages: [{ role: "user", content: "Hi" }],
});

Full script: examples/proxy_node_example.mjs.

curl

curl http://127.0.0.1:8787/proxy/openai/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-User-Id: user_42" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Hi"}]}'

Go

client := openai.NewClient(
  option.WithAPIKey(os.Getenv("OPENAI_API_KEY")),
  option.WithBaseURL("http://127.0.0.1:8787/proxy/openai/v1"),
  option.WithHeader("X-User-Id", "user_42"),
)

Docker sidecar

docker compose -f examples/docker-compose.sidecar.yml up --build

App containers only need OPENAI_BASE_URL / ANTHROPIC_BASE_URL pointing at http://metering-proxy:8787/proxy/....

Or wrap a local process:

python -m agent_metering run --start-proxy -- python your_app.py

Python-only shortcut (optional)

Same venv install auto-patches OpenAI / Anthropic SDKs (no base URL change):

pip install llm-agent-metering
# run your Python app — no import required

Opt out: AGENT_METERING_AUTO=0. Demo: python examples/auto_instrument_example.py.

Also detects FastAPI/Flask/Django request users and OpenAI user= when present.

Who this is for / not for

For: Any stack that can set an LLM HTTP base URL (or env) and needs per-customer / per-user / per-feature spend.

Not for: Full tracing/evals (Langfuse), or replacing multi-provider gateways you already run (LiteLLM / Portkey) unless you put this proxy in front.

Why

Flat API rate limits do not protect margin. Agent workloads are open-ended: tool loops and long contexts can burn tokens quietly. Metering per customer/feature surfaces that before margin disappears.

Project layout

Primary (any language): proxy.py, providers/, cli.py
Python convenience: autoload.py (.pth), instrument.py, user_detect.py, frameworks.py
Shared: config.py, core.py, storage.py, context.py

Tests

pytest

Releasing

Maintainers publish to PyPI via GitHub Actions Trusted Publishing (no API token in secrets).

1. One-time on pypi.org: Publishing → Pending publisher
   • Project: llm-agent-metering
   • Owner: prantakhandaker
   • Repository: agent_metering
   • Workflow: publish.yml
   • Environment: leave empty
2. Bump version in pyproject.toml to match the release tag.
3. Tag and release:

# version in pyproject.toml must match the tag
git tag v0.3.1
git push origin v0.3.1
# then GitHub → Releases → Draft release from that tag → Publish

Publishing workflow: .github/workflows/publish.yml.

License

MIT