Get a free NVIDIA API key for a coding agent
NVIDIA hands out API keys for its hosted open models at no charge: sign in at build.nvidia.com, generate a key, and you get a string that starts with nvapi-. Keep it in a .env file or your agent’s own config, never in a memory file, and a terminal coding agent such as MinniCode runs on it with requests going straight from your computer to NVIDIA.
Updated 2026-11-07
Get the key#
- Open build.nvidia.com and sign in. A personal account is enough; there is nothing to pay.
- Open build.nvidia.com/settings/api-keys and generate a key. It starts with
nvapi-. Copy it now: the page does not show it again. - Pick a model from the catalog. The free inference endpoints host NVIDIA’s own Nemotron family and other open models, such as DeepSeek; the list changes, so read it on the site.
The hosted endpoints speak the same chat API as OpenAI’s, at https://integrate.api.nvidia.com/v1, so any client that takes a base URL and a key works. The free tier has a request limit NVIDIA sets per account; for a coding agent used by one person it is the normal way to run.
Where to keep it#
- In a
.envfile in the project, asNVIDIA_API_KEY=nvapi-..., with.envin.gitignore. Most SDKs and agents read it from there or from the environment. - In the agent’s own config, when it has one. MinniCode asks for the key the first time it runs and keeps it on your computer.
- Never in a memory file. A key in
CLAUDE.md,AGENTS.mdor an auto memory note is sent with every turn, to every model you use, and lands in every transcript. MinniMemoryMCP’sdoctorflags a credential-shaped string in a memory file (MM008) andapplyrefuses to compile it. - Never in a
.env.examplethat is committed. Put the variable name there with an empty value.
NVIDIA_API_KEY=nvapi-your-key-hereIf a key does leak, revoke it on the same settings page and generate a new one.
Run a coding agent on it#
MinniCode is a terminal coding agent built for open models. On first run it asks which provider you want and opens the NVIDIA key page for you; paste the key and pick a model. Requests go from your computer to NVIDIA; Minni’s servers are never in the path and never see the key.
# Windows, in PowerShell
irm https://minniai.com/install | iex
# macOS and Linux
curl -fsSL https://minniai.com/install | shThen type a task in your project folder. Every edit is shown as a diff first; commands, deletions and pushes need a yes. On the same tasks, a warm MinniCode session wrote 5,286 tokens to the model’s cache against 8,273 for Claude Code, 36% less; the method is on the evidence page.
What a free model can and cannot do, OpenRouter as the other free source, and running a model on your own computer are in How to run a coding agent on free, open models. The commands are in the MinniCode docs.