Axis: Agent Surface

What this is

Axis C in the three-axis API model. Canary's most novel axis — and the one with no GK Software equivalent. The AI-native surface where AI agents (Claude Code, Cowork, in-store AI assistants, third-party agentic tooling) are first-class consumers. Three sub-surfaces: MCP tool registry, SSE event streams, outbound webhook publishers.

Purpose

When the consumer is an AI agent rather than a human or a traditional integration, the contract shape changes. Tools have to be machine-described (MCP). Events have to stream so agents can react in real time (SSE). Outbound notifications need typed schemas so agents can subscribe without scraping (webhook publishers). Axis C is what makes Canary AI-native rather than retrofitted-for-AI.

Structure

MCP tool registry

Each agent-facing service exposes an MCP server with named tools, JWT-gated, tenant-scoped:

MCP Server	Tools	Tier mix
`canary-compliance`	7	Reference + Stream
`canary-raas`	7	Reference + Stream
`canary-fox`	TBD	Reference + Stream
`canary-chirp`	TBD	Reference + Change-feed
`canary-owl`	TBD	Reference
`canary-store-brain`	5	Reference + Stream
`canary-ops`	6	Reference + Stream

SSE streams

Server-Sent Events for live event consumption by AI agents:

Stream	Service
`/alerts/sse`	alert
`/transactions/sse`	tsp
`/inventory/sse`	inventory-as-a-service
`/analytics/sse`	analytics
`/ops-dashboard/sse`	ops-dashboard
`/store-brain/sse`	store-brain

Webhook publishers

Outbound notifications from Canary to merchant-configured destinations on alert creation, case escalation, detection rule firing, daily digests, reconciliation summaries.

Direction

Outbound — Canary serves AI consumers.

Why no GK equivalent

GK's Function namespace dispatches generic functions for in-POS plug-ins; that's RPC-flavored, not AI-native. MCP tool surfaces with typed input/output schemas, capability gates via canary-store-brain permission resolution, and AI-readable descriptions are a different model entirely. This is Canary's AI-native differentiator — the surface that makes the platform leapfrog the enterprise POS playbook rather than catch up to it.

Consumers

Claude Code in development sessions
Cowork in operator-facing workflows
In-store AI assistants
Third-party agentic tooling
Founder ALX/ALXjr instances

Permission gating

Every MCP tool call passes through canary-store-brain for permission resolution. The session-scoped policy decides which tools are allowed for which agent at which location. Without permission gating, the MCP surface is too dangerous to expose.

Sources

canary-store-brain SDD (session governance + permission resolution)
canary-compliance SDD (7-tool MCP reference)
canary-raas SDD (7-tool MCP reference)
microservice-architecture.md MCP surface tables

Anti-pattern

Don't expose MCP tools without store-brain permission gating. Don't publish events on SSE streams without tenant scoping at subscribe time. Don't push outbound webhooks to user-supplied URLs without HMAC signing and replay protection.