Set up contextual rules

Contextual rules judge relationships that syntax alone cannot settle. They are off by default and run only when project configuration or the command line enables them.

OpenRouter

OpenRouter is the shortest hosted setup. Put the model choice in pyproject.toml.

[tool.mcmr.contextual]
backend = "openrouter"
model = "deepseek/deepseek-v4-flash-0731"
reasoning_effort = "medium"

Provide the key to the process, then verify the resolved backend before spending tokens.

export OPENROUTER_API_KEY="..."
mcmr backends .
mcmr check . --contextual

MCMR reads the environment. It does not load a .env file itself. Keep secrets outside pyproject.toml and version control.

Local command backends

The codex and claude backends call an already installed and authenticated command line tool. Set binary only when its executable has a different name or path.

[tool.mcmr.contextual]
backend = "codex"
model = "gpt-5.6-terra"
reasoning_effort = "medium"

Both command backends run isolated, schema-constrained, single-turn processes. They do not receive write access from MCMR.

Local GLiNER2

Use gliner2 when model weights are already provisioned locally.

[tool.mcmr.contextual]
backend = "gliner2"
model = "fastino/gliner2-base-v1"
model_path = "/models/gliner2-base-v1"

The path must be an existing model directory. GLiNER2 does not use reasoning effort or hosted API credentials.

Turn the lane off

--contextual and --no-contextual override project configuration for one run. External evidence is separate. A contextual rule that reads DataHub needs both --contextual and --external.

Continue with Contextual tuning before increasing token budgets.