Configuration

Per-call behaviour is configured on @callm(...), shield(...) or complete(...). Process-wide settings use callm.configure:

import callm

callm.configure(
    home="~/.callm",               # data directory: callm.db, pricing.json
    storage="sqlite",              # "sqlite" | "memory" | storage object
    cache_store=None,              # separate cache backend, e.g. RedisCacheStore(...)
    telemetry=True,                # persist call records
    otel=False,                    # OpenTelemetry spans
    enabled=True,                  # False: decorated functions run untouched
    default_retries=2,             # retries for @callm() without retry=
    assumed_output_tokens=1024,    # cost estimate when max_tokens is not set
    clients={},                    # provider name -> SDK client for fallbacks / complete()
    async_clients={},
    on_call=[],                    # callables receiving every CallRecord
)

Environment variables

Variable	Default	Effect
CALLM_HOME	~/.callm	data directory
CALLM_STORAGE	sqlite	sqlite or memory
CALLM_TELEMETRY	1	0 disables persisting call records
CALLM_OTEL	0	1 enables OpenTelemetry spans
CALLM_DISABLED	0	1 turns callm into a no-op (kill switch)

Provider SDKs read their own variables: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY / GEMINI_API_KEY.

Tests and CI

# conftest.py
import callm

def pytest_configure():
    callm.configure(storage="memory", telemetry=False)

Or set CALLM_DISABLED=1 to run your test suite without callm in the loop.

Logging

callm logs to the callm logger (with a NullHandler attached by default). Warnings include skipped fallbacks, flagged injections, cache or telemetry failures and unknown prices; retries are logged at INFO.

import logging
logging.getLogger("callm").setLevel(logging.INFO)