Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
/**
* R3.7 — non-disruptive allowlist signal.
*
* If a Host's configured model is not in the operator allowlist for its
* provider, emit a structured WARN + a metric. mcp-host does NOT reject the
* model — hard enforcement is control-api/WRC. While the allowlist is
* unavailable (degraded-explicit) the check is skipped: with no allowlist to
* compare against, only the Host-configured model is treated as permitted.
*
* Extracted from main.ts so the pure decision is unit-testable without loading
* the runtime module.
*/
import { llmModelNotAllowedTotal } from './allowlistMetrics'
import type { AllowedModelEntry } from './configStore'
/** The slice of ConfigStore this check reads — lets tests pass a stub. */
export interface AllowlistView {
allowlistAvailable(): boolean
allowedModels(): ReadonlyMap<string, AllowedModelEntry[]>
}
/**
* Signal (WARN + metric) when `model` is outside the allowlist for `provider`.
* Returns true iff a violation was signalled. No-op when the model is allowed,
* the allowlist is unavailable, or the model config is absent.
*/
export function signalHostModelAllowlist(
view: AllowlistView,
model: { provider: string; name: string } | undefined
): boolean {
if (!model?.provider || !model.name) return false
if (!view.allowlistAvailable()) return false
const entries = view.allowedModels().get(model.provider) ?? []
if (entries.some(e => e.model === model.name)) return false
llmModelNotAllowedTotal.labels({ provider: model.provider, model: model.name }).inc()
console.warn(
JSON.stringify({
level: 'warn',
event: 'llm_model_not_allowed',
provider: model.provider,
model: model.name,
message:
'Host-configured model is not in the operator allowlist; not enforced by mcp-host (control-api/WRC gate applies to new selections)',
})
)
return true
}