Skip to content

Free LLM API keys for the hive (checked 2026-09-27)

Free tiers let the hive's free-agent workers run without paying. Each provider limits requests or tokens per day. Keys from different providers add up, because each provider counts its own limits.

Where the keys go

  • 24/7 work runs on the owner's Mac. Put the keys in ~/.hive/agents.env there, one NAME=value per line. The Mac's hive-loop.sh picks them up on its next start.
  • GitHub Actions is only for occasional manual runs (for example hive-cloud.yml while the Mac is off, or ai-accounts.yml). Never schedule the hive on GitHub:
  • frequent automated pushes got the owner's previous account flagged as spam;
  • GitHub's terms forbid using hosted runners for anything unrelated to the repository's own project;
  • the free tiers' daily caps would run out anyway.
  • Never commit a key. Keys go only in agents.env on the Mac, or in repository secrets.

The providers, best value first

Limits change often. The numbers below are those listed by the sources at the end on 2026-09-27; the provider's own console shows the live numbers.

# Provider Get the key Free limits (as listed) Env var OpenAI-compatible base URL
1 Cerebras https://cloud.cerebras.ai Reported 1M tokens/day, 10–30 RPM CEREBRAS_API_KEY https://api.cerebras.ai/v1
2 Google AI Studio (Gemini) https://aistudio.google.com/app/apikey Flash models: about 15 RPM and up to about 1,500 requests/day, by model. Pro left the free tier in April 2026 GEMINI_API_KEY https://generativelanguage.googleapis.com/v1beta/openai
3 Groq https://console.groq.com/keys 30 RPM, 250–1,000 requests/day by model GROQ_API_KEY https://api.groq.com/openai/v1
4 NVIDIA NIM https://build.nvidia.com 40 RPM (NVIDIA Developer Program sign-up) NVIDIA_API_KEY https://integrate.api.nvidia.com/v1
5 Mistral https://console.mistral.ai/api-keys Free monthly credits, no card MISTRAL_API_KEY https://api.mistral.ai/v1
6 OpenRouter https://openrouter.ai/keys Free models: 20 RPM, 50 requests/day, or 1,000/day after a one-time $10 credit purchase OPENROUTER_API_KEY https://openrouter.ai/api/v1
7 GitHub Models https://github.com/marketplace/models 150 requests/day (small models), 50/day (large). Uses the workflow's own GITHUB_TOKEN GITHUB_TOKEN https://models.github.ai/inference
8 Cloudflare Workers AI https://dash.cloudflare.com/profile/api-tokens 10,000 "neurons"/day CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID https://api.cloudflare.com/client/v4/accounts/<account id>/ai/v1
9 Ollama Cloud https://ollama.com/settings/keys Free session and weekly limits (not published) OLLAMA_API_KEY https://ollama.com/v1
10 SambaNova Cloud https://cloud.sambanova.ai Small daily limits (sources differ: about 20 RPM, 200K tokens/day) SAMBANOVA_API_KEY https://api.sambanova.ai/v1
11 Cohere https://dashboard.cohere.com/api-keys Trial key: 1,000 calls/month COHERE_API_KEY see Cohere's "Compatibility API" docs
12 Hugging Face Inference Providers https://huggingface.co/settings/tokens Small monthly credit HF_TOKEN https://router.huggingface.co/v1
13 LLM7.io https://token.llm7.io About 60 requests/hour without a key; more with a free token LLM7_API_KEY https://api.llm7.io/v1
14 Kilo Code gateway https://app.kilo.ai/profile About 200 requests/hour KILO_API_KEY https://api.kilo.ai/api/gateway
15 Aion Labs https://www.aionlabs.ai/app/api-keys/ 15 RPM, 20K tokens/day AION_API_KEY https://api.aionlabs.ai/v1
16 OVHcloud AI Endpoints https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/ 2 RPM without sign-up none https://oai.endpoints.kepler.ai.cloud.ovh.net/v1
17 Z AI (Zhipu GLM) https://open.bigmodel.cn/usercenter/apikeys Permanent free Flash models; sign-up may need a Chinese phone number ZAI_API_KEY https://open.bigmodel.cn/api/paas/v4
18 ModelScope https://modelscope.cn/my/myaccesstoken About 2,000 requests/day; needs Chinese phone verification MODELSCOPE_API_KEY https://api-inference.modelscope.cn/v1

Command-line agents on the Mac (use the owner's own login, no API key):

  • Gemini CLI: 1,000 requests/day and 60 requests/minute with a personal Google account.
  • Qwen Code: its free OAuth tier closed on 15 April 2026, so don't rely on it.

Free machines that can run the hive 24/7

The workers are light: plain Python that calls LLM APIs, so almost any always-on machine will do. Connect every machine to the Mac's local hub (~/.hive/hub.git) over Tailscale, which is free for personal use (https://login.tailscale.com/start), so no port is opened to the internet.

Option Cost What you get Watch out for
Oracle Cloud Always Free (https://signup.cloud.oracle.com), home region Mumbai or Hyderabad ₹0; a card is needed only for identity checks Ampere A1 Arm VM. Since June 2026: 2 cores and 12 GB RAM, free for ever • The home region may have no free A1 capacity: retry, or pick the other Indian region.
• An instance whose CPU, network and memory all stay under 20% for 7 days can be reclaimed.
• Pay-As-You-Go accounts reportedly keep 4 cores and 24 GB free, but then set a budget alert, because anything beyond the free shapes is billed
Old Android phone + Termux (Termux from F-Droid: https://f-droid.org/packages/com.termux/) ₹0, no card Python on a phone that stays on • Run termux-wake-lock.
• Set Termux's battery use to "Unrestricted".
• Keep it on the charger and Wi-Fi
Google Cloud e2-micro (https://cloud.google.com/free) ₹0 within the free tier; card needed 1 small VM (1 GB RAM), US regions only Only us-west1, us-central1 and us-east1 are free

Not for 24/7: - GitHub Actions: terms and spam risk, see above. - Google Colab and Kaggle: time-limited sessions; Kaggle is for training only. - GitHub Codespaces: stops when idle. - Hugging Face Spaces: the free CPU tier sleeps after 48 hours without visitors.

Tips from the lists and forums

  1. Spread the load across providers. Each counts its own limits. free_agent.py already moves to the next provider on a 429 or a missing key.
  2. One account per provider. Opening several accounts at the same provider to multiply the free quota breaks almost every provider's terms, and the keys get banned. Different providers are fine.
  3. Small, fast models for bulk work, big ones for judging. Flash, 8B and 20B models are enough for drafting and checking links; keep the large models for the JEVX judge.
  4. OpenRouter's one-time $10 raises its free-model limit from 50 to 1,000 requests a day. It is the cheapest upgrade if one is ever wanted.
  5. Free tiers may use prompts for training. Never send personal data. The hive only sends public exam content.
  6. Model names retire often. When a provider answers 404 for a model, update the model list rather than the code.

Sources