CNXT AI Cost Guardrails

CNXT AI Cost Guardrails gives site administrators a local control plane for plugins that use the native WordPress AI Client.

The plugin records content-minimized operational metadata around native AI Client calls, estimates token-billed cost from rates you control, and can prevent later prompts after recorded monthly usage reaches a configured threshold.

Features

• Sitewide monthly limits for estimated USD, AI call attempts, and total tokens.

• Per-source monthly limits for plugins, mu-plugins, themes, core, and unknown callers.

• Monitoring mode that marks calls which would have been prevented without blocking them.

• Enforcement mode that uses the official pre-flight AI Client prevention filter.

• Atomic request reservations to reduce concurrent overrun of request limits.

• Conservative handling of missing token or price data.

• Emergency stop that prevents new native AI Client prompts immediately.

• Content-free audit rows for prompts blocked by configured limits, fail-closed safety checks, or the emergency stop.

• Provider/model token-rate editor with historical estimate recalculation.

• Usage by source, recent activity, retention controls, and formula-safe monthly CSV export.

• No external service, account, telemetry, remote assets, license check, or paid feature lock.

All functionality included in this plugin is available without payment, registration, a license key, a quota, or a trial period.

Data minimization

CNXT AI Cost Guardrails stores only the time, inferred calling source, provider ID, model ID, capability, token counts, estimated cost, duration, coarse request context, and policy status.

It does not intentionally read or store prompt or response content, direct user identifiers, IP addresses, request URLs, API keys, or connector credentials. It briefly uses object identities, not message text, to correlate lifecycle events within one PHP request. All plugin data remains in the site’s own WordPress database.

The plugin makes no outbound network requests and never starts an AI request.

Important limits

• Coverage is limited to calls that pass through the native WordPress AI Client lifecycle. Direct provider SDK or HTTP calls are outside this plugin.

• Enforcement and the emergency stop are designed for the recommended wp_ai_client_prompt() wrapper. Lower-level SDK calls may bypass the prevention filter.

• If a provider result event never arrives, the plugin retains a content-free row with unknown token and cost values so request limits do not silently fail open.

• The AI Client does not expose a caller ID. Source attribution is inferred from the synchronous PHP stack. Indirect calls may appear as core or unknown.

• Enforcement reserves one request under a MySQL or MariaDB lock before generation. A single call can still exceed a token or USD threshold.

• When token or price information is missing, enforcement is conservative. A configured request limit can serve as the measurable fallback.

• USD values are estimates, not invoices. Cached input, batch pricing, media pricing, free tiers, discounts, taxes, and provider-specific rules may differ.

• Keep provider-side billing caps enabled as the final backstop.

Privacy

The plugin stores content-minimized operational metadata locally for site administrators and does not contact a third party. Operational timestamps and source/context labels may still be sensitive in some environments, so site owners should choose an appropriate retention period. Site owners should also evaluate their AI providers and the plugins initiating AI requests, because those components may process content under separate terms.

Upgrading from a previous slug

Deactivate the previous version before activating CNXT AI Cost Guardrails. Do not run both versions at the same time. Activation migrates existing local settings and renames the legacy events table for each site as that site loads.

Existing limits, emergency-stop state, token rates, and usage history are retained. If legacy storage cannot be migrated safely, native AI Client prompts are paused and an administrator notice is displayed until initialization succeeds.

Custom integrations must use the new cnxt_aicg_ filter prefix and CNXTAICostGuardrails namespace.