The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the LoomaScale Ads listing page.
Self-hostable MCP servers that let ChatGPT or Claude run real ad accounts — with spend guardrails they cannot talk their way past.
MIT licensed. Extracted from a production service that has run live campaigns from inside ChatGPT since 2026.
| What it does | Docs | |
|---|---|---|
| google-ads-mcp | 42 tools over the Google Ads API: performance, keyword research, campaign and Performance Max creation, budget and bidding changes | README |
Meta Ads is next.
No JDK, no Maven, no source build:
Then open http://localhost:8080/connections and connect an ad account.
Set MCP_DEFAULT_MAX_DAILY_BUDGET_CENTS low before pointing this at a live account. It is
the last thing between a model and your card.
Full prerequisites — the Google Ads developer token in particular, which needs an application and takes days to approve — are in the server README.
Worth reading before running anything, because money is involved:
mcp-core is the reusable half: MCP's JSON-RPC transport, a tool
registry SPI, and a minimal OAuth 2.1 authorization server with dynamic client
registration, PKCE and refresh-token rotation. If you want ChatGPT or Claude to call your
tools over an authenticated endpoint, that is the part to depend on — it knows nothing
about advertising beyond a few shared abstractions the platform servers reuse.
One repository, one version. A change spanning the core and a platform is one commit and one build, which is why this is not three repositories:
google-ads/server resolves mcp-core at ${project.version} inside the reactor, so a
fresh clone builds with nothing published anywhere:
Credentials are encrypted at rest with AES-256-GCM. Authorization codes and refresh tokens are stored hashed and are single-use; reusing a rotated refresh token revokes the whole family. Model-supplied image URLs pass an SSRF policy before anything is fetched. Tool arguments are never logged — only argument names.
Put TLS in front of any of this. Report vulnerabilities privately: see SECURITY.md.
These servers are extracted from LoomaScale, an AI operator for Google Ads: you talk to ChatGPT or Claude, and it reads your account, finds waste, and makes the changes. This repository is the engine that runs it — not a demo or a cut-down version. The tools here are the same ones that have created and managed live campaigns since 2026, guardrails included.
What stays in the product rather than here is the commercial half: accounts and billing,
usage metering, the multi-user dashboard, and the connect flows for people who do not want
to run a server. mcp-core exposes those as replaceable interfaces — QuotaPolicy,
ResourceOwnerAuthenticator, AdsConnectionStore — precisely so the open-source build has
working defaults and LoomaScale can substitute its own.
ai.loomascale.com is the hosted version. Same tools, no Docker, no developer token application, no OAuth client to register. It starts with a free Google Ads waste audit — the assistant reads your account and reports what is being wasted, before you pay anything. Plans and limits are on the pricing page.
Self-hosting is a genuine alternative, not a teaser: everything needed to run these servers against your own accounts is in this repository under the MIT licence, with no feature held back and no phone-home.
The connection guides are written for the hosted product but the client-side steps are identical against your own server — the same consent screen, the same connector setup:
Other things extracted from the same codebase live at ai.loomascale.com/oss.
See CONTRIBUTING.md — the secrets section first.
Built and used in production by LoomaScale. MIT licensed.