llmf

Free, keyless LLM chat in one call.

llmf (llmfree) talks to public, no-key OpenAI-compatible endpoints, picks a live one, and falls back automatically when something fails. Zero third-party dependencies — stdlib only. Works on desktop, server, and Termux.

import llmf

print(llmf.chat("Hello — who are you?"))

No API keys. No provider setup. Same return type every time.

Install

pip install -U llmf

Requires Python ≥ 3.8.

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

Quick start

import llmf

# Auto provider + model
llmf.chat("Summarize the CAP theorem in two sentences.")

# Intent shortcuts: fast | best | code | vision | reasoning
llmf.chat("Write a Python fibonacci function", model="code")

# Fuzzy or explicit model ids
llmf.chat("hi", model="llama")
llmf.chat("hi", model="ovh/Meta-Llama-3.3-70B-Instruct")

# Try models in order
llmf.chat("hi", model=["best", "fast"])

# Options
llmf.chat(
    "Return a JSON object with keys name and age",
    system="You only output valid JSON",
    max_tokens=300,
    temperature=0.2,
    json=True,
)

# Images (path, URL, or bytes)
llmf.chat("Describe this image", images=["photo.jpg"], model="vision")

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

chat() returns a Reply: behaves like a str, plus metadata (provider, model, attempts, …).

CLI

llmf "Your prompt here"
llmf chat "Explain recursion" --model best
llmf models              # built-in catalog
llmf models --live       # probe providers (network)
llmf providers
llmf health
llmf version

Local OpenAI-compatible proxy (stdlib HTTP server):

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

Defaults to 127.0.0.1 only. Use --host 0.0.0.0 only with --api-key. Use --cors if a browser extension must call it.

API surface

Environment (optional):

llmf.configure(timeout=30, exclude=["blockrun"])
llmf.configure(proxy="http://127.0.0.1:8080")
llmf.register("lab", "http://localhost:11434/v1", ["llama3"], front=True)

How fallback works

1. Prefer endpoints that are fast and clean in practice.
2. On failure, cool down that target (e.g. longer after HTTP 429).
3. Strip thinking / stats noise; treat empty “success” bodies as failures.
4. If every target fails → llmf.AllProvidersFailed with per-provider .errors.

Built-in providers (availability changes over time): ChatJimmy, OVH, FreeAI, LLM7, Kilo/OpenRouter gateway, BlockRun, Pollinations, ChatAI, and others registered in the package. Always verify with llmf health / llmf models --live.

Safety notes

• These are public, unofficial, free endpoints. They can rate-limit, change, or disappear without notice.
• Do not send secrets, passwords, or personal data — prompts go to third-party servers.
• discover() contacts external catalogs; treat it as opt-in.
• Default local server binds to localhost and disables CORS on purpose.

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 — see LICENSE.

PyPI: pypi.org/project/llmf · Source: github.com/MoonLite-Br/llmfree