OpenAI Ads MCP
Enables management of OpenAI Ads campaigns, ad groups, ads, and insights through MCP, with approval-gated live mutations for safe write operations.
README
OpenAI Ads MCP
A Model Context Protocol server for OpenAI's ChatGPT Ads platform — read campaigns/ad groups/ads, pull insights, draft optimization plans, and apply approval-gated live mutations. Built for use from Claude, OpenClaw, Hermes, or any MCP-compatible agent.
[!IMPORTANT] You need OpenAI Ads API access to actually run this against live data. The server talks to
https://api.ads.openai.com/v1with a Bearer key. As of mid-2026, OpenAI's ChatGPT Ads is an invite/beta program — there is no public self-serve Ads API, only a web Ads Manager and a measurement pixel/Conversions API. Without a key for that endpoint, the network-backed tools fail fast with aMissing OPENAI_ADS_<ACCOUNT>_API_KEYerror before making any network call.So treat this repo as a reference / template MCP: a working blueprint for wrapping OpenAI Ads as an agent tool, ready to use the moment you have API access. The draft/report/validation tools that don't hit the network work standalone, and the test suite runs fully offline against a mocked client.
Safety model
- API keys come from environment variables only — never hardcoded, never in source. A key pasted into a chat or log should be considered compromised; rotate it.
- Read / report / draft tools are available by default and are non-destructive.
- Live mutations are double-gated. A write executes only when both are true:
OPENAI_ADS_ENABLE_WRITES=1, and- a matching approval artifact exists at
tasks/openai-ads/approvals/<id>.jsoninside this repo (shape below) whose declared budget/countries/destructiveness bound the operation.
- Every mutation is audit-logged to
tasks/openai-ads/audit.jsonl(gitignored).
Approval/audit paths resolve relative to this package's directory by default;
override with OPENAI_ADS_REPO_ROOT, OPENAI_ADS_APPROVAL_DIR, or
OPENAI_ADS_AUDIT_LOG.
Quick start
First, clone and install (all three clients need this). Requires
Bun — curl -fsSL https://bun.sh/install | bash.
git clone https://github.com/adityash8/openai-ads-mcp.git
cd openai-ads-mcp
bun install
One env var is all you need: OPENAI_ADS_PRIMARY_API_KEY. No key yet? Skip it —
the server starts fine and the four offline tools work; only the network-backed
tools need a key. Then pick your client below.
Claude Code
From inside the cloned directory, one command ($PWD expands to the absolute
path at add-time):
claude mcp add --env OPENAI_ADS_PRIMARY_API_KEY=YOUR_KEY --transport stdio openai-ads -- bun "$PWD/index.ts"
This registers it for the current project. Add --scope user (before the
openai-ads name) to make it available in every project. Verify with
claude mcp list, then use it in a session: "list my OpenAI Ads campaigns."
Claude Desktop
Edit the config file (Settings → Developer → Edit Config, or directly):
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json
Add the server (use the absolute path to index.ts), then fully quit and
reopen Claude Desktop:
{
"mcpServers": {
"openai-ads": {
"command": "bun",
"args": ["/absolute/path/to/openai-ads-mcp/index.ts"],
"env": {
"OPENAI_ADS_PRIMARY_API_KEY": "YOUR_KEY"
}
}
}
}
The tools appear under the 🔌 / tools menu once the server connects.
ChatGPT
[!WARNING] ChatGPT is more involved than Claude. ChatGPT's custom connectors only accept remote HTTPS MCP servers (SSE / streaming HTTP) — it cannot launch a local stdio server. So you bridge this server to HTTP and expose it over a tunnel. Requires a paid plan (Plus, Pro, Business, Enterprise, or Edu); Developer Mode is not on Free.
-
Bridge stdio → SSE with supergateway:
OPENAI_ADS_PRIMARY_API_KEY=YOUR_KEY \ npx -y supergateway --stdio "bun index.ts" --port 8000 --ssePath /sse -
Expose it over HTTPS with a tunnel (ChatGPT must reach it from the internet) — e.g. cloudflared or ngrok:
cloudflared tunnel --url http://localhost:8000 # prints an https://… URL -
Add the connector in ChatGPT: Settings → Apps → Advanced → Developer mode (enable) → Create app. Set the MCP server URL to your tunnel's
https://…/sseendpoint. It lands under "Drafts," then is callable in chat.
[!CAUTION] A tunnel puts this server on the public internet. Keep
OPENAI_ADS_ENABLE_WRITES=0(the default) so it stays read-only, put auth on the tunnel (supergateway--header "Authorization: Bearer …", or a Cloudflare Access policy), and shut the tunnel down when you're done. Treat the URL as a secret.
Configuration details
- Multiple ad accounts: each account is an uppercase account key; its
key is read from
OPENAI_ADS_<ACCOUNT_KEY>_API_KEY(e.g. addOPENAI_ADS_BRANDX_API_KEYand passaccount_key: "BRANDX"in tool calls). The default account key isPRIMARY; change it withOPENAI_ADS_DEFAULT_ACCOUNT. - Writes are off by default. Set
OPENAI_ADS_ENABLE_WRITES=1and supply an approval artifact (below) to enable mutations. - Pass env vars through your MCP client config (the
envblock /--envflag), as shown above. A local.envworks too, but only when running under Bun from this directory —node dist/index.jsand MCP clients launched from other working directories won't read it. - Prefer a Node runtime?
bun run buildonce, then point your client atnode /absolute/path/to/openai-ads-mcp/dist/index.js(Node 18+). Bun is still required for build/test — plain Node can't run the TypeScript source.
Tools
| Tool | Network | Purpose |
|---|---|---|
openai_ads_get_account |
✓ | Fetch the resolved ad account |
openai_ads_list_campaigns |
✓ | List campaigns |
openai_ads_list_ad_groups |
✓ | List ad groups |
openai_ads_list_ads |
✓ | List ads |
openai_ads_get_insights |
✓ | Account / campaign / ad-group / ad insights |
openai_ads_optimization_report |
✓ | Summarize insights into an optimization read |
openai_ads_upload_image |
✓ | Upload a creative image |
openai_ads_draft_campaign_bundle |
— | Build a campaign→ad-group→ads operation set (no writes) |
openai_ads_draft_ad_update |
— | Draft an ad-update operation (no writes) |
openai_ads_validate_conversion_event |
— | Validate a conversion event against the supported schema |
openai_ads_apply_approved_plan |
✓ | Execute an approved, gated set of mutations |
Approval artifact shape
operations must contain at least one entry. Valid operation types:
create_campaign, update_campaign, campaign_state, create_ad_group,
update_ad_group, ad_group_state, create_ad, update_ad, ad_state.
State operations take { "id": "...", "action": "activate" | "pause" | "archive" }.
{
"id": "example-approval",
"account_key": "PRIMARY",
"approved_by": "you",
"approved_at": "2026-06-02T00:00:00.000Z",
"max_budget_micros": 10000000,
"approved_countries": ["US"],
"allow_destructive": false,
"operations": [
{ "type": "ad_state", "id": "ad_123", "action": "pause" }
]
}
Conversion event shape
openai_ads_validate_conversion_event expects:
{
"type": "order_created",
"timestamp_ms": 1764547200000,
"data": {}
}
type must be one of: appointment_scheduled, checkout_started,
contents_viewed, custom, items_added, lead_created, order_created,
page_viewed, registration_completed, subscription_created,
trial_started. id, timestamp_ms, and metadata are optional.
Develop
bun test # offline — uses a mocked Ads client
bun run build # bundle index.ts -> dist/ (Node target)
Live read-only probe
After setting a real key, verify connectivity without changing anything:
bun run probe:live -- --account PRIMARY
Checks the ad account, campaign listing, and account insights. For insights,
time_granularity is none or daily (all / day are accepted aliases and
normalized). Metric arrays encode as fields[]=impressions&fields[]=clicks.
License
MIT © 2026 Aditya Sheth. See LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。