AdCP Signals
Adaptor.
Reference implementation for the AdCP Signals & Measurement working group.
Verified in the AAO agent registry ↗ Running this agent feeds work back upstream — see the protocol contributions (merged schema fields, conformance-suite fixes, open RFCs).
- Status Production · live demo
- License MIT — open source
- Spec AdCP 3.1.0 (GA)
- Runtime Cloudflare Workers
- Role Reference seller implementation
- Author Founding member · AdCP S&M WG
The fragmentation tax
finally has a settlement.
For two decades the ad-tech industry has lived with a fragmentation tax — OpenRTB for bidding, AdCOM for creative, GPP for privacy, OpenDirect for guaranteed buys. Each one a real standard. Each one its own island. Each one requiring custom bridges. As AI buyer agents enter the stack, the bridges become the bottleneck.
The Ad Context Protocol (AdCP) is the open standard designed to reduce that tax. Built on top of Anthropic's Model Context Protocol, it gives AI agents a single, secure way to discover and transact with advertising platforms — audience activation, curation, and media buy. One spec, many platforms.
The signals adaptor is the reference seller-side implementation of the protocol's data provider role. It serves AdCP spec 3.1.0 (additive over 3.0, so a 3.0 integration keeps working untouched), ships as MIT-licensed code, and runs as a live demo other adopters can probe.
One spec, many platforms.
The bridges were the bottleneck.
AdCP design thesis
A production worker.
The full provider surface.
A production Cloudflare Worker that implements the full AdCP signals provider surface — eight capability surfaces, every payload schema-validated against a vendored copy of the AdCP corpus, and a conformance harness running against the official storyboard scenarios on every push.
- get_signals
Signal discovery
Discover signals across 14 verticals and 5 signal types — exposed to any AdCP-enabled buyer agent.
- activate_signal
Signal activation
Turn a discovered signal into bookable inventory inside the seller catalog.
- query_signals_nl
Natural-language query
Translate buyer intent into a structured signal lookup — no agent prompt-engineering required.
- list_creative_formats · get_products
Creative + product surface
Expose the seller's products and creative formats so buyer agents can plan media buys.
- check_governance
Governance check
Trust + compliance gate before activation — applied as a pre-condition, not an audit.
- create_media_buy · get_operation_status
Media-buy primitive
Idempotent buy creation + status tracking. Idempotency-key handling for safe retries.
- /.well-known/adagents.json
Discovery endpoint
Publisher-discovery endpoint per AdCP convention — buyer agents find the provider by URL pattern.
- X-AdCP-Spec-Version
Spec versioning
Response header pinned to the latest released patch — buyer agents pin compatibility.
Input. Process. Output.
A buyer-side agent issues an MCP request; the worker validates against the spec corpus, looks up signals, translates between MCP and AdCP, and ships a spec-compliant response back to the seller catalog.
Built for the edge.
Probe it yourself.
- MIT open source
GitHub repo
Source + history + issues. Pull requests welcome.
- Production
Live demo
Running adaptor — call it, inspect headers, probe the catalog.
- Endpoint
Discovery JSON
Publisher-discovery endpoint per AdCP convention.
- Essay
Related writing
"AdCP: The Open Standard for Agentic Advertising" — thesis behind the code.
- AgenticAdvertising.org
Standards working group
AdCP Signals & Measurement WG — founding member.
A founding member's
contribution.
AdCP is a project of AgenticAdvertising.org, a neutral consortium advancing open standards for AI-powered advertising. Founding members include Optable, PubMatic, Scope3, Swivel, and Triton Digital, with a supporting roster of AccuWeather, Butler/Till, Classify, Raptive, The Weather Company, Newton Research, Bidcliq, and Samba TV.
I serve as a founding member and contribute to the Signals & Measurement working group. The signals adaptor is one of the reference seller implementations the working group uses to anchor protocol decisions.
The open question in front of the working group now is dimensional. Today’s signal types carry two dimensions of a moment — who and what. The where and the when have no native signal in the stack, and the embeddings that encode them (Google’s PDFM on the spatial axis; hour-of-week as the newer frontier) live in a different geometry from the text embeddings the current signals ride on. Similarity across two geometries is undefined until a spec defines it — which makes bridging them a standards question, not an integration. The Context Economy lays out the argument.
It's also registered on the public agent registry at agenticadvertising.org as a No Fluff Advisory signals provider.