llmf

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

llmf (llmfree) is a tiny Python library and CLI that talks to public, no-API-key OpenAI-compatible endpoints. It picks a working provider, falls back when one fails, and needs zero third-party dependencies (stdlib only). Runs on Linux, macOS, Windows, and Termux.

import llmf

print(llmf.chat("Explain the CAP theorem in two sentences."))

No signup. No keys. Same Reply / OpenAI-shaped objects every time.

Package info (pip show llmf)

pip show llmf
pip show -f llmf    # list installed files

Install

pip install -U llmf

Verify:

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

Quick start

Simple API

import llmf

llmf.chat("Hello!")
llmf.chat("Write fibonacci in Python", model="code")
llmf.chat("hi", model="fast")          # aliases: fast | best | code | vision | reasoning
llmf.chat("hi", model="llama")         # fuzzy name
llmf.chat("hi", model="ovh/Meta-Llama-3.3-70B-Instruct")
llmf.chat("hi", model=["best", "fast"])

llmf.chat(
    "Return JSON with keys name and age",
    system="Only valid JSON",
    max_tokens=256,
    temperature=0.2,
    json=True,
)

llmf.chat("Describe this", images=["photo.jpg"], model="vision")

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

chat() returns a Reply: acts like a str, plus provider/model/attempt metadata.

OpenAI-shaped API (0.3+)

from llmf import LLMF

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

# Streaming
for event in llm.chat.completions.create(model="fast", messages=[...], stream=True):
    ...

Module helpers still work: llmf.chat(...), llmf.models(), llmf.providers(), llmf.health().

llmf.models.list()
llmf.models.check()       # doctor: real call per model (cached ~24h)
llmf.models.resolve("fast")
llmf.providers.list()
llmf.health.check()

CLI

llmf "Your question"                 # one-shot ask
llmf ask "..." -m best --stream
llmf chat                            # interactive (/reset, /exit)
llmf models                          # catalog
llmf models list
llmf models resolve fast
llmf models check                    # live probe + health cache
llmf doctor [--all] [--include-limited]
llmf providers
llmf health
llmf serve [--port 8000] [--api-key SECRET]
llmf version

Useful flags: -m/--model, -p/--provider, -s/--system, --image, --max-tokens, -t/--temperature, --stream, --json, --routing ranked|adaptive, --race N, --proxy URL.

Local OpenAI-compatible server (default 127.0.0.1, no CORS):

llmf serve --port 8000
curl http://127.0.0.1:8000/v1/models

Bind 0.0.0.0 only together with --api-key.

Configuration

Isolated client (no shared state):

from llmf import Client
c = Client(timeout=20, routing="adaptive")
c.chat("hi")

How it works

1. Build a fallback plan from aliases, explicit ids, and provider health.
2. Prefer fast/clean targets; cool down failures (longer on HTTP 429).
3. Optional routing="adaptive" uses latency EWMA; race=N hedges parallel requests.
4. doctor / models.check() results are cached (~24h) to demote dead models.
5. Strip thinking/stats noise; treat empty “HTTP 200” bodies as failures.
6. If everything fails → llmf.AllProvidersFailed with .errors per provider.

Public free endpoints change often — use llmf health and llmf doctor to see what still works.

Safety

• Endpoints are unofficial and free. They may rate-limit, change, or vanish.
• Do not send secrets, passwords, or personal data to third-party servers.
• discover() contacts external catalogs; treat as opt-in.
• Default llmf serve listens on localhost only.

Project links

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/*

Supported Python: 3.8–3.13.

License

MIT © 2026 Html / MoonLite-Br