Default providers
FreeGent ships with the providers below. One key is enough to start; each extra key adds models to fall back on when a free tier runs out. Any other OpenAI-compatible endpoint can be added as a custom model (see Local models).
Free-tier limits are set by the providers and change often, so check each provider's page for current numbers. When a model answers 429, FreeGent puts it on a cooldown and moves to the next one in your priority list.
| Provider | Variable in the credentials file | Notes |
|---|---|---|
| Google Gemini | GEMINI_API_KEY (also GOOGLE_API_KEY) | Free tier from AI Studio |
| OpenRouter | OPENROUTER_API_KEY | Many :free models |
| NVIDIA NIM | NVIDIA_API_KEY | Free API credits at build.nvidia.com |
| Kilo | KILO_API_KEY | :free models; needs a key (keyless access ended in 2026-10) |
| Nous Portal | NOUSPORTAL_API_KEY (also NOUS_API_KEY) | |
| TokenHarbor | TOKENHARBOR_API_KEY | Weekly quota |
| Vercel AI Gateway | VERCEL_API_KEY | Selected models at $0 per token |
| MoonMath | MOONMATH_API_KEY | Free plan: daily allowance on DeepSeek V4.1 Flash and GLM-5.3 Flash; key required. Paid plans meter usage. |
| Requesty | REQUESTY_API_KEY | A few $0 models among 800+ priced ones; they work on a $0 balance, about 2 requests at a time. See models. |
| Groq | GROQ_API_KEY | Rate-limited free tier |
| OpenAI | OPENAI_API_KEY | Paid. Also the key slot for any other OpenAI-compatible endpoint you add |
| Hugging Face | HF_API_KEY (also HUGGINGFACE_API_KEY) | Image generation fallback |
| Ollama, vLLM, LM Studio, any OpenAI-compatible server | none needed | See Local models |
Optional keys for tools:
| Tool | Variable |
|---|---|
| Web search (Tavily) | TAVILY_API_KEY |
| Web search (Brave) | BRAVE_API_KEY |
| GitHub search (raises 60/hr to 5,000/hr) | GITHUB_TOKEN |
| Stack Overflow search | STACKEXCHANGE_API_KEY |
Web search in Auto mode tries Tavily, then Brave, then Wikipedia. Image generation uses Pollinations.ai without a key and falls back to Hugging Face when a key is set.
Keys and variables
Where a key goes depends on how you run FreeGent:
| You run | Put the key | Notes |
|---|---|---|
| freegent.ai | Settings → Models → API Credentials | Stored in your browser |
| Local dev server | ~/.config/freegent/credentials, or Settings | File keys stay on the server |
| TUI | credentials file, shell variables, or --api-key | See TUI |
| CLI | credentials file, shell variables, or --api-key | See CLI |
API Credentials also has a paste box and an Extract from file button. Paste a block of keys (or pick a .env-style file) and FreeGent fills each provider's field by recognising the key's prefix. It asks before replacing a key you already saved.
Keys are never included in profiles or in chat logs you send.
MoonMath (its Zro API) uses https://zro.moonmath.ai/v1/chat/completions. Its free daily allowance resets at 00:00 UTC; the provider does not publish a token or request limit. Add a key in Settings or set MOONMATH_API_KEY in the credentials file. CLI example: freegent --provider moonmath --model deepseek-v4.1-flash. For the paste box, use MOONMATH_API_KEY=sk-...; the bare sk- prefix is shared by other providers. The catalog includes only the two models advertised for the Free plan, whose availability was checked against the pricing and API docs on 2026-11-01. Authenticated completions and streamed tool-call round trips were verified for both models on that date. The daily Free allowance is small: one model-ladder conversation with DeepSeek V4.1 Flash used it up after about 2M tokens (37 requests, 81K context), and the reply named the reset ("It resets in 21h 42m").
Check for updates queries authenticated GET https://zro.moonmath.ai/v1/models, deduplicates zro/ aliases, and checks the two known free-plan model IDs. The API does not mark free-plan eligibility, so newly added free-plan models need the allowlist updated; the updater does not automatically assume every accessible model is free.
MoonMath's two Flash routes passed image-reading checks (random six-digit codes) and appear in image-model routing.
Requesty
Requesty uses https://router.requesty.ai/v1/chat/completions with provider/model IDs. Set REQUESTY_API_KEY in the credentials file, or paste the key (rqsty-...) in Settings. CLI example: freegent --provider requesty --model novita/ling-3.1-flash. Its public GET /v1/models lists over 800 models with per-token prices. On 2026-11-01, 13 of them cost $0 for input and output. Those work on an organization with a $0 balance; priced ones answer 402 "Your organization's balance is too low to run this request". Streamed tool calls were checked on the free routes the catalog includes: Ling 3.1 Flash (Novita), Nemotron 3 Ultra, Nemotron 3.5 Lightning, Muse Glimmer 30B, Nemotron 3 Nano Omni, Gemma 4 31B and Leanstral 1.5 (a Lean 4 proof model). Muse Glimmer and Nano Omni read an image with a random code. Gemma 4 is listed with vision but answered 400 to images, so it is text-only. Three $0 routes are listed but gone upstream (Ling 3.0 Tiny and Nemotron 3 Super: 404; Nemotron 3 Nano 30B: 410) and are blacklisted. Two Laguna routes have no tool support, and Nemotron 3.5 Content Safety is a classifier. Requesty's metadata marks the NVIDIA-served routes as possibly used for training, with 30-day retention; each catalog note says so.
Check for updates reads the public model list and proposes $0 routes with tool support, carrying over their context size and image and reasoning flags. reasoning_effort is sent on every Requesty model. Off sends none, which Requesty documents for all models.
Requesty has no per-minute limit. It caps the requests an organization has in flight: the balance rounded up to the next dollar, times 5 for the lowest tiers (in-flight limits). At a $0 balance, two of eight parallel requests went through on 2026-11-01. The others got 429 "Your organization reached its limit of concurrent requests", with no Retry-After. FreeGent retries after 5 seconds and holds every Requesty model for those seconds, since the limit is shared. With "one request per endpoint" on (the default), each model sends one request at a time. A key's monthly spend limit answers 412 "Reached monthly spend limit for API key"; every Requesty model then waits an hour and is checked again, because the end of the cycle is not reported. Free models also share an undocumented budget: after about 15M tokens in two hours (model ladder, 2026-11-01) every $0 route answered 429 "You've hit the rate limit for free models", with no Retry-After or reset time. FreeGent then cools all Requesty models together for an hour at a time: the limit still held more than an hour after it was hit. Upstream rate limits (429 "Too many requests", origin provider) cool only that model, as for any provider. Requesty allows browser calls (Access-Control-Allow-Origin: *).
Quota errors for MoonMath
MoonMath uses the shared retry engine, preserving HTTP status and provider error text. The MoonMath API reference documents 402 for exhausted spend quotas or plan budgets and 429 for rate limits. A quota error cools both MoonMath models until the reset named in the message, else until 00:00 UTC: the pricing page says only that the Free plan's allowance "resets every day", and the account console showed "Today's Free usage: 89% left · resets in 1h 2m" at 22:58 UTC on 2026-11-01. The console shows the allowance only as a percentage; it held about 2M tokens. A key can also carry its own spend limit (Key controls on the API keys page). Reaching it returns 402 key_quota_exceeded ("This API key has reached its spend limit") even while Free usage is left; FreeGent re-checks that hourly, since the owner can raise the limit at any time. The Free plan also runs at most two requests at once per key (x-ratelimit-api_key-limit-max_parallel_requests: 2). A third gets 429 "The Free plan runs a limited number of requests at once" with Retry-After: 60; a slot frees as soon as one request finishes, so FreeGent retries that model after 5 seconds and does not cool the other one. With "one request per endpoint" on (the default), two MoonMath models at most are in flight.
Removed: APInex
APInex (api.apinex.bond) was added and removed on 2026-11-01. Its free/ routes need a daily check-in on its site, behind a CAPTCHA, before they answer, and the free allowance (20M weighted tokens a day, about 5 to 10M real tokens) ran out within one model-ladder conversation at 153K context (GPT-6 Luna: 12 of 21 microtasks). Saved APInex keys, models and priority-list entries are dropped on load.