Run CodeCouncil 100% free with an NVIDIA API key (full setup guide)

[adigo-pro]

Run CodeCouncil 100% free with an NVIDIA API key

CodeCouncil's default model is NVIDIA-hosted Nemotron 3 Super — chosen in our bake-off for zero false positives — and NVIDIA hosts it (plus 100+ other open models) behind a free API key. No credit card, no pi login, no trial clock. This guide walks through the whole thing end to end, including how to switch models and how to run council mode entirely free.

1. Get the free key (≈2 minutes)

1. Go to build.nvidia.com and click Login / Join (top right). Any email works — signing up enrolls you in the free NVIDIA Developer Program. No credit card, no company info.
2. Once signed in, generate a key either way:
   • build.nvidia.com/settings/api-keys → Generate API Key, or
   • open any model page (e.g. Nemotron 3 Super) and click Get API Key.
3. Copy the key — it starts with nvapi-. One key works for every model in the catalog.

What the free tier gives you (as of mid-2026): new Developer Program accounts get ~1,000 free inference credits (1 credit ≈ 1 API call) and a rate limit of ~40 requests/minute. That's far more headroom than CodeCouncil needs — the critic only calls the model when your code actually changed, with a floor between calls, so a typical coding session uses a handful of calls per hour. (NVIDIA's own reference: docs.api.nvidia.com.)

2. Install CodeCouncil

curl -fsSL https://raw.githubusercontent.com/adigo-pro/CodeCouncil/main/install.sh | sh

That checks Python 3.10+, clones to ~/.codecouncil/app, puts codecouncil on your PATH, wires up pi (the model runtime) if npm is available, and scaffolds ~/.codecouncil/env for the next step. (install.sh is ~102 audited lines — git clone + python3 -m codecouncil works identically if you prefer manual.)

3. Add the key

echo 'NVIDIA_API_KEY=nvapi-...' >> ~/.codecouncil/env

~/.codecouncil/env lives outside any repo, so the key can never end up in a commit — CodeCouncil loads it automatically on every model call. (Alternatively, run /keys inside a running council for a guided, hidden-input setup.)

4. Run it

codecouncil /path/to/repo-you-code-in     # defaults to the current directory

With NVIDIA_API_KEY present and no model configured, CodeCouncil defaults to NVIDIA-hosted Nemotron automatically — nothing else to configure. Code with Claude Code in another terminal; findings arrive in your terminal, on the dashboard (localhost:4700), and inside your coding agent's own context.

Switching models (still free)

The running council is interactive — switch the model in place, no restart needed on your side:

/model nvidia-nim/deepseek-ai/deepseek-v4-pro

/model <provider/model-id> sets and persists the primary model (it restarts just the critic). Every one of these is served by the same free NVIDIA key:

/model value	What it is
nvidia-nim/nvidia/nemotron-3-super-120b-a12b	Default. Precision anchor from the bake-off — 0 false positives
nvidia-nim/nvidia/nemotron-3-ultra-550b-a55b	Bigger Nemotron, slower
nvidia-nim/nvidia/nemotron-nano-3-30b-a3b	Fastest/lightest
nvidia-nim/deepseek-ai/deepseek-v4-pro	Strong decorrelated candidate (different model family)
nvidia-nim/deepseek-ai/deepseek-v4-flash	Faster DeepSeek
nvidia-nim/z-ai/glm-5.2	Another decorrelated family

You can also set it before launch with COUNCIL_MODEL=nvidia-nim/... codecouncil /path/to/repo — precedence is CLI flag > env var > saved config. /status shows what's currently resolved and /config shows where each value came from.

Model IDs must be NVIDIA's full catalog string (nvidia-nim/nvidia/nemotron-3-super-120b-a12b, not nvidia-nim/nemotron-3-super) — NVIDIA's API 404s on anything shorter.

Bonus: council mode, also free

Council mode adds a second, independently-trained model (the prober) that reviews the exact same evidence — a prober-only finding is only delivered after CodeCouncil reproduces it for real. The README's example prober uses OpenRouter, but you can run a fully-free council off the one NVIDIA key by picking a prober from a different model family than the primary:

/prober nvidia-nim/deepseek-ai/deepseek-v4-pro

(/prober off turns it back off.) Different training lineage is the point — the whole premise of CodeCouncil is a second pair of differently-trained eyes, and Nemotron + DeepSeek don't share blind spots.

Troubleshooting

Symptom	Fix
401 / auth errors	Key typo or truncated — re-check the nvapi- line in ~/.codecouncil/env (no quotes needed, no spaces around =)
404 on model calls	The model ID isn't the full catalog string — use the exact nvidia-nim/... values from the table above
Critic silent, no verdicts	Run /status in the council to see daemons, last verdict, and key state; /verbose unmutes idle-beat chatter
Rate-limit errors	The free tier is ~40 req/min — CodeCouncil stays well under this; if you hit it, something else is sharing the key
Ran out of free credits	Sign back into build.nvidia.com — Developer Program accounts can request more free credits, or switch providers with /keys + /model

---

Questions welcome below — and if you get it running, the 60-second start in the README is the short version of this guide.