RRetailAgentOS
Developers & Platform Architects

The technical gateway to RetailAgentOS

An open specification and a tested reference engine, not a walled product. This page links to the primary sources instead of duplicating them.

610/610 automated tests passing · 93.8% line coverage (rules engine scope) — see the scorecard

Delivery surfaces

UCP and projections remain the canonical interoperable representations. Native browser WebMCP now adds controlled local actions; a generalized remote/server MCP remains a designed future transport, not a hosted service.

UCP + projections

Manifest, specs, feeds, and UI all derive from canonical merchant policy objects.

Browser-local WebMCP

Discover controlled tools and prepare allowed actions without reimplementing decisions.

Remote MCP, future

Inspect the browser package and the remote/server architecture; the latter is documented, not shipped.

What RetailAgentOS is, technically

An open specification (RAOS) plus a tested reference engine that evaluates a retailer’s eligibility, pricing, inventory, fulfilment and quote rules against a buyer’s context, and returns a deterministic, explainable DecisionRecord.

Relationship to UCP

UCP defines the discovery and transport rails commerce agents use to talk to a store. RetailAgentOS is the reasoning layer that sits behind those rails — it decides what an agent should be told, UCP carries the message.

Reference-engine overview

Published as @retailagentos/engine — pure, deterministic evaluation functions with no I/O, no wall-clock reads, and no randomness. Same inputs always produce the same DecisionRecord.

See the architecture→

Decision pipeline overview

A staged, fault-isolated pipeline runs each extension in priority order, folding reasons into one ordered record. A throwing evaluator degrades safely instead of crashing the pipeline.

See the pipeline model→

Extension model

Each capability — eligibility, pricing, inventory, fulfilment, quote integrity, trust — is a namespaced extension (com.os.retailagent.shopping.*) with a pure evaluate function, negotiated per merchant via the discovery manifest.

See the extension registry→

Decision records and reasons

Every decision carries a uniform reason vocabulary (code, severity, source, requirements) and renders into audience-specific traces — merchant ops, buyer-facing, or raw developer JSON.

Known limitations

  • Native browser WebMCP tools ship at /webmcp-showcase; the generalized remote/server MCP is designed, not shipped.
  • Cryptographic signing is simulated by locked decision, labeled TRUST_SIMULATED wherever it applies.
  • No agent authentication, rate limiting, persistence, or multi-tenant production infrastructure.
  • Promotion stacking, loyalty and restricted-goods enforcement are designed but not implemented.
See the full public scorecard →