webGCP — Web Graph Context Protocol

An open protocol for AI agents to request typed knowledge bundles with provenance from any compliant server. Read the v0.1 specification →

v0.1 working draft. The spec host (webgcp.org) is the first publicly-deployed host honoring §5.1.1 DNS-layer SVCB discovery; the spec is self-conformant against L0 fixtures WGCP-L0-001/002/003. The L1 reference instance (MyParallel — the BigQuery property-graph substrate) is in-build; deployed-vertical L0/L1/L2 references are held until cognition services land and process real work. L2–L4 are design targets. v0.2 candidate sketches the Commerce-Layer surface (NOT yet v0.2 final). Do not cite for production conformance claims until v0.2 publishes. Full maturity statement →

What this is

Agents in production today behave like this: an agent gets a task, decides it needs information, searches, retrieves, evaluates, decides it needs more, searches again. Most of the token and latency budget is spent before the model attempts the task. The dominant fix has been better retrieval — better embeddings, better rerankers, bigger context windows. None of these address the underlying problem, which is that agents are missing a typed contract for the knowledge they need.

webGCP defines that contract. A small set of typed objects, served at predictable URLs, with provenance on every field:

Manifests describe what a knowledge base is for, when it applies, what inputs it needs, what authority its contents carry.
Bundle contracts describe what an agent needs as a typed schema.
KL-Query is the declarative wire format an agent uses to request a bundle.
Typed responses carry the bundle with per-field _meta — source, timestamp, authority, confidence.
Discovery uses /.well-known/webgcp to make any compliant server reachable without prior coordination.

The protocol is implementation-neutral. A compliant server can use BigQuery, Snowflake, Postgres+pgvector, a graph database, or a hand-rolled store. The protocol constrains the shapes agents see, not the stores operators use.

What this is not

webGCP composes with, does not replace:

MCP (Anthropic / AAIF) — tool calling; webGCP adds the typed knowledge layer above it.
WebMCP (W3C CG) — browser-side tool exposure; webGCP adds server-side persistent knowledge.
x402 (payment) — settlement; webGCP carries x402 references in receipts when present.
AP2 (mandate) — authorization; webGCP consumes AP2 assertions in the invoker identity.

An operator that runs both this protocol (the "Context layer") and a payment / routing / marketplace stack (a "Commerce layer" like the one at the-ai-news.web.app) is the expected pattern — see spec §0.5 for the layering framing.

For wire-level pairings with the upstream supporting protocols (MCP, WebMCP, SEP-1649, AP2, x402 V2, A2A, DNSAID), see /composition/.

The reference substrate

The Network's reference implementation runs on a BigQuery property-graph substrate at webgcporg.webgcp_network.* — the source-of-truth for skills, agents, roles, tasks per §5.7 + §6.3. Two-tier distribution per the substrate's public/private split:

Anonymous tier — signed content-addressed artifacts; integrity by signature, not access control. Anyone reads, no auth required.
Authenticated tier — BigQuery Analytics Hub linked datasets; subscription is the access control. No cross-project IAM grants.

Today the substrate is queryable via the live cognition-engine-mcp endpoint (in production since 2026-05-17). Direct anonymous-artifact distribution and Analytics Hub listing are the in-build design targets per the substrate's hard-shard topology. This is the live source-of-truth for The Network; not a parallel "demo" substrate.

How to read this

If you are a…	Start at
Engineer who wants to build a conformant server	§4 Conformance Levels, §5 Normative Requirements, §9 Conformance Tests
Agent author who wants to consume webGCP	§3 Mental Model, §6 Wire Format, §7 Bundle Contracts
Skeptic	§1.4 Known Weaknesses, §12 Open Questions
Standards reviewer	§1 Why, §2 Design Principles, §11 Governance

Quick links

/spec/v0.1/	The specification (HTML) — frozen normative
/spec/v0.2-candidate/	v0.2 candidate (Commerce-Layer surface; NOT yet v0.2 final)
/composition/	Wire-level pairings with MCP / WebMCP / SEP-1649 / AP2 / x402 / A2A / DNSAID
/request-for-development/	Implementer's MVP guide — build sequence, gates, common pitfalls
/dns-setup/svcb-at-squarespace/	Registrar-specific: how to add the §5.1.1 SVCB record at Squarespace Domains (worked example + fallbacks)
/schemas/manifest/v0.1.json	Manifest schema (JSON Schema 2020-12)
/schemas/bundle/v0.1.json	Bundle contract schema
/schemas/entity/v0.1.json	Entity (skill row) schema
/schemas/receipt/v0.1.json	Signed-receipt schema
/conformance/v0.1/	WGCP-L0-001 + WGCP-L0-002 fixtures (runnable)
/.well-known/webgcp	Machine-readable spec-host descriptor
/implementations/	Conformant implementations (1 live)
/governance/	Governance status
/contact/	Get in touch — collaboration intake + spec questions

Build something

The shortest path from "I just read this spec" to "I have a conformant server":

Read /request-for-development/ — the implementer's MVP guide. Construction sequence, gates, common pitfalls, suggested project skeleton, stack-substitution table. ~8–16 hours of active engineering to L0.
Skim §4 (conformance levels) + §5 + §6 in the spec — the normative MUSTs and the wire format.
Curl the live L0 reference endpoint to see typed responses with _meta for yourself.
Build to WGCP-L0-001, L0-002, and L0-003 — when all three pass, you're L0-conformant.
File issues at the contact in /.well-known/webgcp.

The first non-author who ships a conformant server completes the second-implementer test — the gating event for v1.0 promotion of any feature in the spec.