Self-hostable MCP connectors for major advertising platforms.
Use this repo when you want transparent platform integrations, local experimentation,
and infrastructure you control. Use Cesteral Intelligence when your team needs
approvals before spend commits, credential brokering, auditability, and
cross-platform execution from one governed environment.
| |
LinkedIn Ads management and analytics — CRUD, targeting, delivery forecasts, and media uploads
This repo is an open connector layer, not the full product.
Self-host a connector for transparency and direct control of your
credentials, infrastructure, and platform access.
Use Cesteral Intelligence when the workflow needs governed writes, team
approvals, credential brokering, auditability, and cross-platform
coordination — see when you need it below.
Flagship Connectors
Google Ads MCP
Campaign writes, GAQL reporting, bid adjustments, previews, and validation via
Google Ads REST API v25.
Thirteen servers, 290+ tools. Tool counts are the live registered total per
server, including the *_search_tools discovery tool where present.
What Every Server Ships
Every server speaks MCP protocol revisions 2025-03-26, 2025-06-18 and
2025-11-25, the ones @modelcontextprotocol/sdk 1.31.0 implements; the
Streamable HTTP transport rejects any other MCP-Protocol-Version. The current
revision, 2026-07-28, is not supported yet
(#242).
That is not the whole spec. In the MCP conformance suite
(@modelcontextprotocol/conformance 0.1.16) every server passes 18 of the 30
active server scenarios — initialize, ping, logging level and log
notifications, tool list and call, elicitation (form, defaults and enums),
resource list, read and templates, prompt list and get, concurrent SSE streams,
DNS-rebinding protection — and
conformance/expected-failures.yaml
gives the reason for each of the other 12. Not implemented: completions,
resource subscriptions, sampling, progress notifications, image / audio /
embedded-resource content, and SSE stream resumption. Elicitation requests
(including the destructive-action confirmations below) and log notifications
are sent on the calling tools/call request's response stream, so a client
does not need the optional GET stream to receive them. A server bound to a
loopback address (the default outside NODE_ENV=production) accepts only
loopback Host and Origin headers on /mcp; a server on any other address
checks Host only when MCP_ALLOWED_HOSTS is set.
Beyond raw tool calls, every server ships:
MCP Prompts — on-demand, multi-step workflow guidance (campaign launch,
reporting, troubleshooting) so agents don't have to rediscover each
platform's sequencing. Present on all 13 servers.
MCP Resources — structured, addressable context (schemas, field catalogs,
examples, enums) fetched on demand instead of bloating every tool schema.
Present on all 13 servers. DV360 uses these to keep its >1 MB discriminated
unions off the wire (entity-schema://, entity-fields://, entity-examples://).
Tool discovery — a <platform>_search_tools tool that lets an agent
search the server's own catalog by intent instead of paging the full list.
On 10 of the 13 servers (all except the small gads-mcp and sa360-mcp /
reporting-only dbm-mcp).
Server discovery cards — SEP-2127 metadata at
/.well-known/mcp/server-card.json (name, version, transports, auth modes,
capabilities) on every server, in every auth mode.
OAuth resource discovery — in jwt auth mode, the RFC 9728 endpoint at
/.well-known/oauth-protected-resource.
Report CSV spill — large report bodies spill to GCS behind a signed URL
so responses stay bounded. On the six reporting-heavy servers: ttd-mcp,
tiktok-mcp, snapchat-mcp, amazon-dsp-mcp, pinterest-mcp, msads-mcp.
Server
Discovery
Prompts
Resources
CSV spill
gads-mcp
✅
✅
meta-mcp
✅
✅
✅
dv360-mcp
✅
✅
✅
ttd-mcp
✅
✅
✅
✅
linkedin-mcp
✅
✅
✅
tiktok-mcp
✅
✅
✅
✅
cm360-mcp
✅
✅
✅
sa360-mcp
✅
✅
pinterest-mcp
✅
✅
✅
✅
snapchat-mcp
✅
✅
✅
✅
amazon-dsp-mcp
✅
✅
✅
✅
msads-mcp
✅
✅
✅
✅
dbm-mcp
✅
✅
Built for Production
Self-hosting an AI agent that touches live ad spend is a trust problem first
and a capability problem second. Two things make this fleet shippable to
production without hand-rolling guardrails:
Audit-grade observability. Every tool call is captured as append-only
JSONL. Failures additionally capture the full upstream HTTP trail — every
request, every retry, every response — with secrets redacted at the source.
Query it directly in BigQuery for hosted deployments, or pipe stdout to
your existing log stack for self-host.
Read the observability guide.
Destructive-action elicitation gates. 51 destructive tools across the
twelve write-capable servers (dbm-mcp is reporting-only) prompt the user
before deletes, bulk status changes, bid adjustments, budget changes,
conversion uploads, and async Workflows batch jobs. Stdio and clients
without elicitation support fall back to a documented non-interactive
contract. Bulk mutations under 10 items skip the prompt unless they touch a
sensitive field (status / budget / bid).
Verifiable release provenance. Governed tools carry a canonical
SHA-256 definitionHash (from @cesteral/contract-hash) emitted into a
per-package cesteral-manifest.json. Tagged releases publish to npm with
build provenance, signing the manifest transitively inside the tarball, so a
downstream governance system can verify exactly which tool definitions
shipped and promote matching tools to attested trust.
If your security review needs evidence — the redaction list, the field
schema, the upstream capture path, the contract hash — all of it is in this
repository.
pnpm >= 10.0.0 -- corepack enable picks up the pinned version from packageManager (or npm install -g pnpm@10)
Docker (for containerization)
Terraform >= 1.6.0 (for deployment)
1. Clone and install
git clone https://github.com/cesteral/mcp-open-advertising.git
cd mcp-open-advertising
pnpm install
2. Build
pnpm run build
3. Run a server locally
# Start any server using the dev script
./scripts/dev-server.sh gads-mcp # port 3004
./scripts/dev-server.sh meta-mcp # port 3005
./scripts/dev-server.sh dv360-mcp # port 3002
# See each package README for required environment variables
Cesteral uses a GCP-native architecture with thirteen independently deployable Cloud Run MCP services. Each server exposes the MCP protocol directly via HTTPS on Cloud Run.
Direct HTTP transport: No edge gateway layer needed
Independent deployment: Each server can be deployed and scaled separately
Composable: Use one server or all thirteen -- they work independently or together
AI clients connect directly to one or more MCP servers over HTTPS; there is no
shared gateway to provision. Cesteral Intelligence, when used, layers tenancy,
credentials, approvals, and governance above the same fleet.
# Install dependencies
pnpm install
# Build all packages
pnpm run build
# Run tests
pnpm run test
# Type checking
pnpm run typecheck
# Lint
pnpm run lint
Testing MCP Tools
# Use MCP Inspector
npx @modelcontextprotocol/inspector packages/gads-mcp
# Or use curl
curl -X POST http://localhost:3004/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc": "2.0", "method": "tools/list", "id": 1}'