Configuration

MCMR reads [tool.mcmr] from pyproject.toml. Missing configuration uses safe local defaults.

[tool.mcmr]
select = ["*"]
ignore = []

[tool.mcmr.scan]
suffixes = []

select and ignore accept identifiers, prefixes, globs, or callable substrings. suffixes adds source extensions to discovery.

Execution and rule policy

[tool.mcmr.execution]
deterministic = true
contextual = false
external = false

Command flags override these values for one run. Control rules explains selection patterns, lane toggles, exclusions, typed settings, and policy overrides.

Contextual models

[tool.mcmr.contextual]
backend = "openrouter"
model = "deepseek/deepseek-v4-flash-0731"
reasoning_effort = "medium"
timeout_seconds = 180
minimum_confidence = 0.6
batch_size = 32
prompt_token_budget = 128000
max_output_tokens = 32000

Other fields include binary and model_path. Supported backends are gliner2, codex, claude, and openrouter. Start with Set up contextual rules, then use Contextual tuning for request and token limits.

Providers

[tool.mcmr.providers.datahub]
server = "http://localhost:8080"
max_assets = 500
publish_runs = false

Each provider validates its own settings. The DataHub provider can also read DATAHUB_GMS_URL and DATAHUB_GMS_TOKEN. A recorded directory replaces the live server for offline examples.