llmf

Free, keyless LLM access in one call — with automatic multi-provider fallback.

llmf (llmfree) is a stdlib-only Python library and CLI for public, no-API-key OpenAI-compatible endpoints. It picks a live target, falls back on errors, and runs on desktop and Termux.

import llmf

r = llmf.chat("Explain the CAP theorem in two sentences.")
print(r)              # Reply is a str subclass
print(r.source)       # "provider/model"
print(r.provider, r.model, round(r.latency, 3), r.usage)

Package metadata (pip show llmf)

Name: llmf
Version: 0.5.1
Summary: llmf (llmfree): free, keyless LLM chat with automatic multi-provider fallback — stdlib only.
Home-page: https://github.com/MoonLite-Br/llmfree
Author: Html
Author-email: moonlite-br@users.noreply.github.com
License: MIT
Requires-Python: >=3.8
Requires:

pip install -U llmf
pip show llmf

Install

pip install -U llmf
python -c "import llmf; print(llmf.__version__)"
llmf version

Python API

Short form → Reply

import llmf

r = llmf.chat("Hello!")
assert isinstance(r, str)
r.text                 # same as str(r)
r.source               # "provider/model"
r.provider, r.model
r.latency              # seconds
r.usage                # dict or None
r.finish_reason
r.reasoning            # optional
r.attempts             # list of Attempt(provider, model, ok, latency, error)

model= aliases: fast | best | code | vision | reasoning
Also fuzzy names, full provider/id, a list tried in order, or provider="…".

llmf.chat("hi", model="code", max_tokens=256, temperature=0.3)
llmf.chat("hi", model=["best", "fast"])
llmf.chat("Describe this", images=["photo.jpg"], model="vision")  # vision_timeout default 90s
llmf.chat("…", provider="ch-at")       # force provider (works for opt-in too)
llmf.chat("…", provider="nottrack")    # opt-in provider (not in auto chain)

for piece in llmf.stream("Count 1 to 5"):
    print(piece, end="", flush=True)

OpenAI-shaped form

from llmf import LLMF

llm = LLMF()
c = llm.chat.completions.create(
    model="best",
    messages=[{"role": "user", "content": "Hello"}],
)
print(c.choices[0].message.content)
print(c.model, c.provider, c.usage, c.latency, c.attempts)

for chunk in llm.chat.completions.create(
    model="fast",
    messages=[{"role": "user", "content": "Count to 3"}],
    stream=True,
):
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="")

Catalog, doctor, config

llmf.models.list()
llmf.models.list(live=True)                              # chat/free only by default
llmf.models.list(live=True, include_non_chat=True)
llmf.models.resolve("best")
llmf.models.check()                                      # real call per model; ~24h cache
llmf.providers.list()
llmf.health.check()

llmf.configure(timeout=30, vision_timeout=90, routing="adaptive", race=2)
llmf.register("lab", "http://localhost:11434/v1", ["llama3"], front=True)
llmf.discover()                                          # opt-in external catalog

Env: LLMF_TIMEOUT, LLMF_PROXY, LLMF_MODEL, LLMF_ROUTING.
Proxy: http://, https://, socks5://, socks5h://, socks4:// (stdlib).

CLI

usage: llmf [options] "question"              ask once (also reads stdin)
       llmf chat                              interactive chat (/reset, /exit)
       llmf models [list|check|resolve NAME]
       llmf doctor [--all] [--include-non-chat] [--include-limited]
       llmf providers | health
       llmf serve [--port N]
       llmf version

Flags: -m/--model, -p/--provider, -s/--system, -i/--interactive, --image, --max-tokens, -t/--temperature, --stream, --json, -v, --routing ranked|adaptive, --race N, --timeout, --proxy, --live, --all, --include-limited, --include-non-chat, --no-save, --host, --port, --api-key, --cors, --strict-models, --no-fallback.

$ llmf "Say hi in one short sentence."
$ llmf -v "2+2=?"
$ llmf --json "Say hi in one short sentence."
$ llmf models
$ llmf models resolve fast
$ llmf doctor
$ llmf serve --port 8000

What’s new in 0.5.0

• Providers: ch-at (early in chain / best), rewind-ai (Qwen / gpt-4o-mini; sign-in-only marked broken), nottrack (SSE debate adapter — opt-in only, not in automatic fallback; use provider="nottrack").
• ProviderSpec.auto=False: stays out of the auto chain, still usable with provider=.
• BlockRun / Pollinations catalogue tweaks (openai-fast ≡ openai).

Earlier (0.4.x): smarter 429 classes, chat/free-only live catalog by default, dynamic fast tag, vision_timeout, Retry-After scaling — see CHANGELOG.md.

Behaviour notes

• Fallback prefers faster/cleaner targets; failures get cooldowns (rate-limit / capacity / paid messages differ).
• routing="adaptive" uses latency EWMA; race=N hedges parallel requests.
• doctor / models.check() results cached ~24h.
• Total failure → llmf.AllProvidersFailed with .errors.

Safety

• Endpoints are unofficial and free; they may change or vanish.
• Do not send secrets or personal data.
• Opt-in providers (nottrack) and discover() are explicit.
• llmf serve defaults to localhost, no CORS, unless you pass --cors / --api-key.

Development

git clone https://github.com/MoonLite-Br/llmfree.git
cd llmfree
python -m unittest discover -s tests -v
python -m build && twine check dist/*

License

MIT © 2026 Html / MoonLite-Br