Free Tiers Were Already There. freellmpool Makes One Faucet.

The free tiers already exist. Groq, Cerebras, Gemini, OpenRouter, Cloudflare, Hugging Face, Mistral, Cohere and a long list of quieter endpoints each expose some capacity you do not pay for — until the day they do not, or until the daily cap fills, or until the model id you bookmarked goes stale.

What is missing is not another signup. It is one place, on your own machine, where those scraps become a single OpenAI-compatible faucet that coding agents and scripts already know how to drink from.

That is the niche freellmpool claims. On GitHub it is 0xzr/freellmpool: MIT-licensed, Python 3.11+, one main runtime dependency (httpx). We installed the current PyPI release — 0.13.0 — on a Linux box on 1 October 2026. The package reports the same version. The repository sat at 119 stars when we checked the GitHub API the same morning.

The author’s own numbers, repeated on PyPI and the project docs, are sharper than a marketing line: twenty-two cataloged provider groups, 178 enabled chat routes, 431 cataloged chat models, with keyless start when an enabled keyless route is up. freellmpool doctor on our install summarised the live catalog as 22 providers and 178 models. We treat “431 cataloged” as the author’s broader inventory claim; we did not independently recount every disabled or pin-only entry.

Several routes need no key. On a fresh install, doctor listed Pollinations, OVHcloud and Kilo Gateway as keyless, with LLM7 key-optional, and capacity status labelled four of those providers healthy before we added any credentials. Lifetime stats after our short probes showed successful free requests completing. Free tiers remain shared and change without notice — freellmpool’s own limitations section says so, and says plainly it does not rotate accounts or evade quotas.

The shape of the tool matters more than the catalogue size. You can ask once from the CLI, import a Pool in Python, run freellmpool mcp for Model Context Protocol clients, or start freellmpool proxy and point OPENAI_BASE_URL at http://localhost:8080/v1. There is also an experimental Anthropic-compatible path for tools that speak that dialect. Profiles and freellmpool code <agent> print copy-pastable setup for Codex, Claude Code, aider, Cline, Continue, Cursor, Hermes and OpenCode — without silently rewriting third-party config.

Routing is local bookkeeping: least-used-first by default, with optional fast, fair, quality and other modes; cooldown after 429s; per-day counters that reset at UTC midnight on the freellmpool side while upstreams keep their own windows. Quality routing tries to save scarce strong free models for harder prompts. None of that turns a free tier into a frontier model. The docs are blunt: good for drafting, classification, triage and everyday coding — not a substitute for GPT-class reasoning on the hardest problems.

It is also not a privacy layer. Prompts go to the selected upstream. The proxy is meant for local or single-user use and binds to loopback by default; exposing it wants a bearer key. npm publication for the OpenCode plugins was still pending as of the author’s 29 August 2026 note — use their repository-local install path if you need that surface.

Compared with OpenRouter, this is not a hosted aggregator you pay per token. Compared with LiteLLM, you are not primarily bringing paid keys to a universal SDK — you are squeezing providers’ own free tiers, with a keyless first reply when the weather is kind. Other routers exist; freellmpool’s bet is the smallest pip-installable path that still speaks the dialects agents already use.

One hundred and nineteen stars is not obscurity theatre, and it is not a household name either. For anyone already juggling free keys across tabs, the interesting move is quieter: stop teaching every agent a new base URL, and put one local faucet in front of the tiers you already have.