Stacking 34 free AI tiers behind one router yields 7.4 billion tokens a month

The core insight here is arithmetic: each of the 34 AI labs offering a free tier gives you something modest in isolation — a few million tokens a month, a few thousand requests a day. Stack all of them behind a single OpenAI-compatible endpoint and you get roughly 7.4 billion tokens per month of working inference capacity across 635 provider/model endpoints. FreeLLMAPI is the machinery that makes that stacking practical.

The router is the real product. It tracks per-key RPM, RPD, TPM, and TPD counters against each provider's reported ceilings, picks the highest-priority healthy model for each request, and on a 429 or 5xx it cools that key down and retries the next one in your fallback chain. Six routing strategies let you optimize for speed, capability, or reliability. Provider keys are AES-256-GCM encrypted in SQLite and decrypted in memory per request; your apps only ever see a single unified bearer token. The whole thing runs on Node 20+ at around 40 MB RSS at idle.

What makes this more than a thin proxy is the catalog maintenance problem it solves. Free-tier offerings shift constantly — models launch, quotas tighten, wire formats break. The router pulls a signed catalog twice daily and applies changes to your local database, verified against a pinned Ed25519 key before anything is written. Free installs get the monthly snapshot, which lags the live feed by 30 days and currently sits about 303 models behind. The $19/year premium tier closes that gap to same-day.

The surface area is wider than chat. The router exposes image generation, video generation, audio speech and transcription, embeddings, and the Anthropic Messages wire format alongside the standard OpenAI surfaces — so Claude Code and the Anthropic SDK work against your free pool without modification. There's also a fusion mode that fans a prompt out to multiple free models in parallel and uses a judge model to synthesize one answer from the drafts.

For agent builders specifically, the MCP server at /v1/mcp lets agents introspect usable models and provider health mid-session. Setup generators for Claude Code, Codex CLI, Aider, Cline, Continue, and a dozen others configure themselves from your live catalog with a single npx freellmapi setup-* command, backing up existing config before touching anything.

The README is admirably honest about the trade-offs. No frontier models. Variable latency. No SLA. Effective intelligence degrades late in the day as top models hit daily caps, then resets at UTC midnight. The disclaimer is direct: this is for personal experimentation, not production — and the per-provider terms-of-service review it references is a genuinely useful addition that most similar projects skip entirely.

The free-tier aggregation idea isn't new, but the execution here — self-updating signed catalog, per-key quota tracking, six routing strategies, native Anthropic and Gemini wire support, a desktop tray app, and honest documentation of the limitations — is more complete than most attempts at the same problem.