About Free Inference

What is here: every entry gives you free access to LLM inference you can build with. That access comes one of three ways: an API key you call directly, a login to a CLI or tool that carries the model, or inference bundled inside a coding tool or builder. Each entry states which, because a key isn't always the way in.

The test for inclusion is whether you can produce an artifact you keep: code, an app, a website, a design, an agent, a skill. Converse-only free tiers (ChatGPT, Claude.ai, Gemini app, deepseek.com, Grok, Copilot) are excluded on purpose — they don't give you something to build with.

What counts as free: strictly $0. No entry requires a credit card, ever. Expiring trial credits are flagged as such in the cost column rather than presented as a standing free tier.

Free takes four shapes here: rate-limited free (forever, throttled), promo free (limited time), free gateways (OpenRouter, OpenCode Zen), and trial credits (one-time, expires). Verification is mixed and honest about it: rows are checked by a nightly live probe where the provider exposes an API, and by hand against official docs where it doesn't. The date on a row is when we last checked it, not a guarantee it holds tomorrow. Limits move without notice. This is a map, not a contract.

Definitions

RPM — Requests per minute — API calls in any 60-second window.

TPM — Tokens per minute — input + output tokens across all calls in a minute.

RPD — Requests per day — hard daily cap, resets on the provider's clock.

Day tokens — Daily or periodic token/compute quota (some providers meter a token pool per day, month, or one-time credits instead).

Timeout — Max request duration before the provider kills the connection.

Context — Maximum context window available on the free tier.

Agent access

Machine data: data/providers.json — validate against data/schema.json
Agent summary: llms.txt
WebMCP: Browser-native tools registered via navigator.modelContext (search_models, get_provider, list_providers, compare_models)
This page and the repo README are generated from that one file — edit the JSON, run python3 build.py, done.

Contribute

Add or correct a row in data/providers.json with a link to your official limits page and the date it was verified, then open a PR. Unverifiable rows get rejected.