agent-carbon 0.1.0

Carbon emissions tracking for LLM agent frameworks

agent-carbon

Carbon emissions tracking for LLM agent frameworks.

Tokens ✓ · Cost ✓ · CO₂e — until now, missing.

agent-carbon is a drop-in callback that plugs into LangChain, LangGraph, CrewAI, and AutoGen the same way token/cost callbacks do — no code changes to your agent, no network calls, no telemetry.

---

Install

pip install agent-carbon

With optional framework integrations:

pip install "agent-carbon[langgraph]"
pip install "agent-carbon[crewai]"
pip install "agent-carbon[autogen]"

---

3-line usage

from agent_carbon import CarbonCallbackHandler

handler = CarbonCallbackHandler(region="us-east-1")
chain.invoke({"input": "..."}, config={"callbacks": [handler]})
print(handler.receipt)

---

Sample receipt

CarbonReceipt(total_tokens=4823, total_gco2e=0.1802, total_wh=0.3650,
              savings_vs_frontier_gco2e=1.0241, calls=3)

## Carbon Receipt

| Metric                          | Value                                 |
|---------------------------------|---------------------------------------|
| Total tokens                    | 4,823 (3,201 in / 1,622 out)          |
| Total energy                    | 0.3650 Wh                             |
| Total CO₂e                      | 0.1802 gCO₂e                          |
| Savings vs `claude-opus-4-6`    | 1.0241 gCO₂e                          |
| Region                          | us-east-1                             |
| LLM calls                       | 3                                     |
| Estimate                        | True                                  |
| Duration                        | 4.21s                                 |

> *Emissions estimated from token counts × model-specific energy factors
> (Wh/1M tokens) × regional grid intensity (gCO₂e/kWh).
> All values are approximate; see factors_meta.json for sources and caveats.*

### Per-call Breakdown

| # | Model               | In tok | Out tok | gCO₂e  | Wh     |
|---|---------------------|--------|---------|--------|--------|
| 1 | `claude-sonnet-4-6` | 1,234  |   512   | 0.0724 | 0.1465 |
| 2 | `claude-sonnet-4-6` | 1,489  |   601   | 0.0843 | 0.1707 |
| 3 | `claude-sonnet-4-6` |   478  |   509   | 0.0235 | 0.0476 |

---

API reference

CarbonCallbackHandler

handler = CarbonCallbackHandler(
    region="us-east-1",           # optional — cloud region for grid intensity
    frontier_model="claude-opus-4-6",  # optional — savings comparison baseline
)

Pass it as a LangChain callback and read .receipt after the run:

result = chain.invoke(input, config={"callbacks": [handler]})
receipt = handler.receipt          # CarbonReceipt
receipt.to_markdown()              # human-readable string
receipt.to_dict()                  # JSON-serializable dict
handler.reset()                    # clear for next run (keeps region/config)

@track_carbon decorator

For plain functions that use LangChain under the hood:

from agent_carbon import track_carbon

@track_carbon(region="us-west-2")
def run_my_chain(query, config=None):
    return chain.invoke({"input": query}, config=config)

result, receipt = run_my_chain("What is solar energy?")

LangGraph helper

from agent_carbon.integrations.langgraph import invoke_with_carbon

result, receipt = invoke_with_carbon(graph, state, region="us-west-2")

CrewAI helper

from agent_carbon.integrations.crewai import kickoff_with_carbon

result, receipt = kickoff_with_carbon(crew, region="eu-west-1")

AutoGen wrapper

from agent_carbon.integrations.autogen import CarbonModelClient

client = CarbonModelClient(base_openai_client, region="us-east-1")
# pass client to AssistantAgent(model_client=client)
print(client.receipt)

---

Supported models

Model	Provider
gpt-5, gpt-5-mini, gpt-5-nano	OpenAI
claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5	Anthropic
gemini-2.5-pro, gemini-2.5-flash	Google
llama-3.3-70b, llama-3.3-8b	Meta
mistral-large, mistral-small	Mistral

Unknown models emit gco2e=0 and wh=0 with estimated=True rather than raising. Add entries to factors_meta.json and call agent_carbon.factors.FACTORS.update(...) to extend the table at runtime.

---

Region awareness

Pass any major AWS, GCP, or Azure region slug to get grid-adjusted CO₂e:

# Oregon's clean grid (~135 gCO₂/kWh) vs. global avg (~494 gCO₂/kWh)
handler = CarbonCallbackHandler(region="us-west-2")

Omit region to use the IEA 2022 global average (494 gCO₂e/kWh). See src/agent_carbon/regions.py for the full table.

---

Methodology & honesty

Every receipt includes:

• estimate: true — always. No precise published inference energy data exists for any major commercial LLM as of 2026.
• methodology — a one-line disclaimer describing the calculation.
• Per-entry source comments in factors.py and factors_meta.json citing the basis for each estimate.

Energy factors are derived from GPU hardware benchmarks for commercial-scale inference (high-batch-size, tensor-parallel, H100/A100 deployments). Estimates should be treated as order-of-magnitude guidance, not accounting figures.

---

Contributing

Contributions to improve factor accuracy — especially from published energy audits or provider sustainability reports — are very welcome. Update both src/agent_carbon/factors.py and factors_meta.json, cite the source, and open a PR.

---

License

MIT