from llm_localfirst import Router, LocalUnavailable

router = Router.from_env()
try:
    out = router.complete("Redact all PII from this record.",
                          source=customer_record, sensitive=True)
except LocalUnavailable:
    # Local model is down. We did NOT send the record to the cloud. You decide.
    ...

pip install llm-localfirst              # the routing brain — zero provider SDKs
pip install "llm-localfirst[openai]"    # + talk to local Ollama/vLLM/LM Studio (and cloud OpenAI)
pip install "llm-localfirst[anthropic]" # + Claude (the default cloud fallback / reason model)
pip install "llm-localfirst[all]"       # everything (also: mcp, pydantic-ai)

ollama pull qwen2.5:7b          # any OpenAI-compatible local server works
pip install "llm-localfirst[openai,anthropic]"
export ANTHROPIC_API_KEY=sk-ant-...   # only needed for the cloud fallback / reason path

from llm_localfirst import Router, Kind

router = Router.from_env()

# 1) Inspect routing WITHOUT spending a token.
print(router.decide(kind=Kind.BULK))     # -> local  (cheap + private)
print(router.decide(kind="reason"))      # -> cloud  (the hard part)
print(router.decide(sensitive=True))     # -> local  (pinned; never cloud)

# 2) Actually run it. Bulk work prefers local, and falls back to cloud only if local is down.
print(router.complete("Summarize this in one sentence.",
                      source=long_text, kind=Kind.BULK).text)

llm-localfirst doctor                      # show config, the allowlist, and local up/down
llm-localfirst route "summarize this" --kind bulk
llm-localfirst route "redact this" --sensitive    # exits non-zero if local is down (fail-closed)

from pydantic_ai import Agent
from llm_localfirst import Router
from llm_localfirst.integrations.pydantic_ai import attach_worker

router = Router.from_env()
director = Agent("anthropic:claude-haiku-4-5", system_prompt="...")

# Adds a `delegate_to_worker(task, source)` tool that routes to your LOCAL model.
# attach_worker REFUSES a non-local worker, so delegated source text can't leak.
attach_worker(director, router, worker_model="local",
              on_delegate=lambda task, result: ...)  # optional observability hook

pip install "llm-localfirst[mcp]"
llm-localfirst mcp        # serves over stdio

router = Router.from_env()
router.complete("summarise this", source=long_document)

router.ledger.calls("cloud")            # 1
router.ledger.tokens("local")           # Usage(input_tokens=..., output_tokens=...)
router.ledger.snapshot()                # JSON-safe, for logs

from llm_localfirst import Budget, Router

router = Router(..., budget=Budget(max_cloud_tokens=200_000))
...
llm_localfirst.BudgetExceeded: cloud token budget spent: 203_400/200_000 tokens

export LF_PRICES='{"haiku": [0.8, 4.0], "sonnet": [3.0, 15.0], "opus": [15.0, 75.0]}'
export LF_MAX_CLOUD_COST=5.00

llm-localfirst complete "..." --usage    # tally on stderr, completion on stdout
llm-localfirst doctor                    # shows the budget and which models are priced

uv venv && uv pip install -e '.[dev]'
ruff check . && pytest