Google Ads MCP server with read, audit, and guarded write tools — allowlisted accounts, two-phase confirmation, paused by default.
Server has solid naming conventions (verb_noun pattern), comprehensive parameter schemas with types and descriptions, and good safety-focused design. However, output schemas are not documented, error handling lacks recovery guidance, and several tools have descriptions under the 50-200 char LLM-optimized range. Tool descriptions average ~150 chars (acceptable) but lack actionable context for when to use each tool vs. alternatives. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification in code.
Audit a Google Ads account — strictly read-only. Runs a series of checks (account overview, conversion tracking, campaign structure, budget efficiency, etc.) and returns findings with severity levels (critical, warning, info).
Create a Performance Max campaign from a specification. Builds budget, campaign, criteria, asset group, and assets in a single atomic mutate. Everything is created PAUSED.
Create a Google Ads Search campaign from a YAML config — safely. Builds budget, campaign, campaign criteria (geo, language), ad group, keywords, and Responsive Search Ad. Everything is created PAUSED by default.
Create a Standard Shopping campaign with listing group tree and shared negative keyword lists. Builds budget, campaign, ad group, product ad, and listing group criteria. Everything is created PAUSED.
Enable a campaign with two-phase confirmation. Returns a preview and confirmation token; second call with token executes the enable.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., campaign_id from list_campaigns). Audit returns 'findings with severity levels' but structure is not specified.
Error handling lacks recovery guidance. No tool description explains what to do if customer_id is not in allowlist, YAML config is invalid, or API quota is exceeded. Errors will not guide LLM self-correction.
Tool annotations missing. Despite clear risk stratification (READ_ONLY vs WRITE, two-phase confirmation), no tools declare readOnlyHint or destructiveHint. LLMs cannot infer safety properties from names alone.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
List campaigns in a Google Ads account with optional filtering by status.
Pause a campaign with two-phase confirmation. Returns a preview and confirmation token; second call with token executes the pause.
Update a campaign's daily budget with two-phase confirmation. Returns a preview and confirmation token; second call with token executes the change.
Confirmation token pattern not self-documenting. update_campaign_budget, pause_campaign, enable_campaign require two calls with a token, but descriptions do not explain the workflow or what happens if token is omitted/invalid.
Parameter dependencies undocumented. create_pmax_campaign has conditional params (target_cpa only valid if bidding='maximize_conversions', target_roas only if bidding='maximize_conversion_value'), but descriptions do not state these constraints.