Skip to content

v0.2.1

Latest

Choose a tag to compare

@github-actions github-actions released this 19 Oct 15:34
· 2 commits to main since this release

Google Ads MCP v0.2.1

Safe Operator release for Google Ads API v25.

  • Preserves the original 13 read tools and adds server/profile metadata, campaign diagnostics, search-term and asset reports (21 pinned report templates including app_asset_performance, pmax_asset_performance, and geographic_performance), channel-aware asset inventory fallbacks (app_assets and asset_group_assets), conversion goals, recommendations, keyword ideas, and forecasts.
  • Adds 20 typed write previews (including ad_update_preview) and four operation interfaces (46 total MCP tools). No public tool accepts a raw protobuf, mutate payload, or Google service name.
  • Requires runtime, profile, and unchanged 30-day policy approval; enforces exact customer/currency scope, a fingerprint-bound campaign allowlist for enable, and explicit monetary limits.
  • Uses Google validate_only, one-time signed ten-minute receipts, durable SQLite applying intent, no ambiguous-write retries, direct readback, drift invalidation, replay protection, and recovery verification.
  • Creates Search, Performance Max (brand_guidelines_enabled=False), and App campaigns paused, rejects duplicate create matches, and reconciles positive location/language campaign targeting (PRESENCE) on paused or enabled campaigns while preserving negative criteria. Generic criterion updates are limited to positive ad-group keywords; status-only child enables bind authoritative CPC evidence; campaign_bidding_preview supports MANUAL_CPC (Enhanced CPC disabled), TARGET_CPA, MAXIMIZE_CONVERSIONS, MAXIMIZE_CONVERSION_VALUE, and TARGET_ROAS with v25 leaf field masks; ad_update_preview updates status and creative fields on RESPONSIVE_SEARCH_AD and APP_AD ads. Campaign enable supports SEARCH, MULTI_CHANNEL (APP_CAMPAIGN), and PERFORMANCE_MAX with channel-specific readiness checks, and aggregate-budget evidence fails closed on malformed or conflicting rows. Typed forecasts preserve match types and accept at most 20 unique geo IDs. Recommendation actions fail closed when validate_only is unavailable.
  • Adds hidden-input system-keyring onboarding plus policy, operation, profile, and secret-management CLI commands.
  • Keeps billing, account linking, user access and lists, offline jobs, conversion uploads, User Data services, schedulers, hosted transports, and autonomous optimization outside the release boundary.

Publication remains gated on the explicitly risk-accepted exact-head production mutation matrix, production read-only regression, paused Performance Max/App acceptance, bounded Search serving evidence, clean Codex/Claude install, CI/security checks, SBOM, checksums, and provenance. Google recommends a separate test hierarchy; this release does not claim that evidence.