This is not a benchmark or a demo config. This is the exact setup the original Nagual runs on right now, on a 1 vCPU / 4 GB VPS, at a total LLM cost of $0/month. Published so you can reproduce it — and so you can see what "runs on free tiers" means in practice, not in marketing.
Snapshot date: July 15, 2026.
| Provider | Keys | Cost | Notes |
|---|---|---|---|
NVIDIA NIM (build.nvidia.com) |
5 | $0 | The workhorse. Free tier per account; five accounts = five parallel quota pools |
| Google AI Studio | 2 | $0 | Gemini models as the mid-cascade fallback |
OpenRouter (:free models) |
1 | $0 | Last-resort reserve; :free models 429 often — treat as emergency lane only |
Five keys on one provider is the trick that makes free tiers livable: the router spreads every model across all five keys, so one throttled key costs you a slot, not a provider.
Roles: dialog = owner-facing voice, background = the 33 loops' fuel.
Order = cascade priority. Every row exists ×5 NVIDIA keys.
| Model | Role | Why it holds the line |
|---|---|---|
z-ai/glm-5.2 |
dialog | Best free reasoning we found for the owner-facing voice |
mistralai/mistral-small-4-119b-2603 |
dialog, background | Fast, stable, rarely throttled |
minimaxai/minimax-m3 |
dialog, background | Strong long-form generation |
nvidia/nemotron-3-super-120b-a12b |
dialog, background | Reliable mid-weight loop fuel |
deepseek-ai/deepseek-v4-pro |
dialog, background | The deepest free thinker; also powers the Moltbook social loop |
nvidia/nemotron-3-ultra-550b-a55b |
dialog, background | Heavy artillery when quotas allow |
- 33 concurrent loops running 24/7 with zero LLM spend.
- Moltbook karma 4 → 630+ in two weeks — earned autonomously: the agent writes posts, comments, and makes friends on the social network for AI agents (moltbook.com/u/Nagual) with no human in the loop.
- Flat memory: RSS ~3 GiB steady after the backpressure fixes; survives on 4 GB + swap.
- Restart-proof identity: model swaps mid-conversation and process restarts do not reset the persona — memory and goals live on the data volume, not in the model.
- Provider catalogs lie. A model listed in the catalog may 404 on hosted inference
(we lost hours to a "available" model that wasn't). Probe every slot with a real
request before trusting it — the
slot_healer_loopdoes this continuously. - Per-model quotas, not per-provider. One throttled model must not kill the provider for every other model on the same key.
- Strip think-tags. Free reasoning models leak
<think>blocks; the router strips them before anything reaches memory or the owner. - Swap is non-negotiable on 4 GB. The reading swarm spikes; swap absorbs it.
- 429 is a routing event, not an error. Cascade sideways (next key), then down (next model). With 5 keys × 6 models the agent has ~30 lanes before it ever fails.
The skeleton is stronger than its free-tier brain. The meta work — grounding judge, skill forge, orchestrator — is where a strong model changes everything:
- Anthropic (
ANTHROPIC_API_KEY): the call path is already wired in. Claude Sonnet/Opus on the meta role is the intended "full awakening" config. - Kimi (Moonshot), Qwen, GPT, GLM, DeepSeek official APIs: any OpenAI-compatible endpoint slots straight into the router config.
- Subscription-style access (like a Claude Code plan) is the economical sweet spot — pay-per-token frontier APIs on 33 always-on loops get expensive fast. That is exactly why the free-tier cascade exists: the organism must stay alive on $0, and think deeper when you can afford it.
This is the experiment we care about most: the same organism on different brains.
If you run it — on free tiers, on Kimi, on Qwen, on Claude — open an issue with the
my-nagual-run label and show us:
- your key/model mix,
- karma curve or loop stats (
/api/statusgives you everything), - the strangest thing it did on its own initiative.
The most interesting runs get featured here.