NVIDIA Build
This is NVIDIA's hosted developer inference API — no container, no GPU of your own required, keyed per-operator from build.nvidia.com. It's a different product from self-hosted NIM: don't conflate the two.
Why prefer it
Local Aider+Ollama is valuable offline and for tiny swarm edits, but on many laptops and WSL boxes it is latency-heavy (see Local Models and its memory-safe default). NVIDIA Build is OpenAI-compatible, commonly free-tier for catalog models, and typically far faster than CPU/iGPU local inference.
Configure
export NVIDIA_API_KEY=nvapi-...
| Base URL | https://integrate.api.nvidia.com/v1 |
| Auth header | Authorization: Bearer $NVIDIA_API_KEY |
| Example model id | nvidia/llama-3.3-nemotron-super-49b-v1 |
NGC_API_KEY (a container-registry / NIM-pull credential) is accepted as a
fallback if NVIDIA_API_KEY is absent, but it isn't guaranteed to carry
inference scope — the probe below always does a live check, so a wrong-scope
key simply falls back to local rather than succeeding silently.
Check whether it's live
python scripts/nvidia_build_probe.py --probe
Preference rule
| Key present? | Reachable? | Operator override? | Backend |
|---|---|---|---|
| no | — | — | local |
| yes | no | — | local |
| yes | yes | prefer local | local |
| yes | yes | no | build |
An absent or unreachable key never crashes a session — it falls back to local. An explicit operator preference for local always wins, even with a live Build key.
Aider wiring
openai-api-base: https://integrate.api.nvidia.com/v1
model: openai/nvidia/llama-3.3-nemotron-super-49b-v1
export OPENAI_API_KEY=$NVIDIA_API_KEY
A key on one machine is not project law
Same rule as Local Models's multi-clone guidance — keys are per-operator, machine-local secrets, never committed, never assumed shared with a teammate.
Build access and rate limits are NVIDIA's terms and can change without notice; Anchor does not require a paid account for this default path.