Carbon of Compute connector
A read-only connector that gives AI assistants the figures published on carbonofcompute.com: the carbon and water footprint of AI language models per unit of output, and the manufacturing footprint of AI accelerators, servers and datacenter plant.
Endpoint
https://mcp.carbonofcompute.com/mcp
Streamable HTTP. No account or sign-in. Every tool only reads data.
Add it to Claude
- Open Customize → Connectors and add a custom connector by URL.
- Name it Carbon of Compute, paste the endpoint above, and save. Leave the OAuth fields empty.
- In a chat, turn it on from + → Connectors.
Add it to ChatGPT
Once the app is listed, search for Carbon of Compute in the ChatGPT apps directory. Before then, an account with developer mode turned on can add the endpoint above as a custom connector with No authentication.
Example questions
- What is the carbon footprint of 1,000 output tokens from Llama 3.3 70B on the Texas grid?
- Compare DeepSeek R1, Qwen3 235B and gpt-oss-120b per coding task on the EU average grid.
- How much of an H100's manufacturing carbon comes from its HBM memory?
- What is the manufacturing carbon of an 8-GPU H100 server, line by line?
Tools
| Tool | What it returns |
|---|---|
list_models | List the AI language models in the Carbon of Compute catalogue, with provider, open or closed weights, evidence level (measured, derived or estimated), whether it is a reasoning model, and the hardware assumed. Use it to find the model id for get_model_footprint. Optional filters narrow the list. |
get_model_footprint | Return the full-stack carbon (gCO2e) and water (mL) footprint of one AI model, split into operational electricity, hardware manufacturing, facility plant and optional training, with low and high values. Also returns energy per unit and a link to the same view on carbonofcompute.com. Defaults match the site: US average grid, PUE 1.15, 5-year hardware life, 33% busy time, training excluded, per 1,000 output tokens. The low-high span is not complete uncertainty; check deployment differences before using a figure in an emissions inventory. |
rank_models | Rank AI models by total carbon per unit under one set of assumptions, lowest first, as the carbonofcompute.com leaderboard does. Can be limited to a list of models to compare, or filtered by weights, evidence level or provider. Each row has the total, its low-high range and the split into operational, hardware manufacturing, facility and training. |
get_accelerator_manufacturing | Return the cradle-to-gate manufacturing carbon (kgCO2e) and fab water (L) of AI accelerators (NVIDIA H100, H200, B200, A100, AMD MI300X), split into logic die, HBM memory, package and board, with the vendor product carbon footprint where one is published. Also returns the per-GPU-slot values the calculator uses, which add a share of the host server. |
get_server_manufacturing | Return the manufacturing carbon (kgCO2e) of an 8-GPU NVIDIA H100 AI server line by line: CPUs, system memory, network cards, circuit boards, power supplies, chassis, NVSwitch, network fabric share, storage and the eight accelerators. Also returns host-silicon fab water. |
get_facility_manufacturing | Return the datacenter facility factors used by Carbon of Compute: manufacturing carbon of the power chain and cooling plant (a line-item bill of materials anchored to manufacturer EPDs), standby generator fuel, refrigerant leakage, and on-site water use by cooling regime, each per kWh of IT energy. The building shell is excluded. |
list_scenario_options | List the values accepted by the scenario inputs of get_model_footprint and rank_models: grid regions with their carbon intensity, datacenter PUE profiles, cooling regimes, accelerator options, hardware lifetimes, busy-time scenarios, training options and workload presets. |
get_methodology | Return a short summary of how Carbon of Compute estimates AI footprints: the four parts counted, the functional unit, the default assumptions, what is left out, and links to the full method report. Use it to explain or caveat a number. |
Where the numbers come from
The connector carries the data of the current carbonofcompute.com release and runs the same calculation as the site, so a figure from the connector matches the site under the same settings. Each answer includes a link to that view on the site. The method, sources and limits are in the method report. Hardware manufacturing is built on REEL LCI, a life cycle inventory of electronics manufacturing whose foreground inventory is built from public sources; the kgCO₂e totals link it to licensed ecoinvent 3.12 background results.
Release data digest (sha256): 120e1935e2590dbec0521eefd76262b2f87bd35244bea98c1e110cb1493b21d0, read 2026-10-27T05:25:34Z.
Data licence and citation
The figures are licensed under Creative Commons Attribution 4.0 (CC BY 4.0, https://creativecommons.org/licenses/by/4.0/). Cite as: Balsvik, J. The Carbon of Compute. https://carbonofcompute.com/
Support
Questions, errors in a figure, or a model you would like added: info@reellci.com.