Bring your own agent

This is a live preview of the whole architecture, built for Claude Code; other agent harnesses that speak the Model Context Protocol (MCP) may also work. You bring the agent and drive the full lifecycle yourself — author a policy, propose a product, and read back the verdict — against the real Shopify catalog over the Universal Commerce Protocol (UCP), a standard for exposing structured product data.

This is a preview, not the MCP integration surface

Don’t confuse this with the MCP page — that’s the managed, policy-authoring surface (author only). Here, we run the orchestrator and the rest of the infrastructure, with an MCP on top that your agent connects to directly to drive the whole flow: author a policy, propose a product, and read back the verdict. It’s the bring-your-own-agent counterpart to the hosted Son of Anton preview, where delta runs the agent and the UI for you.

Prerequisites

• The Claude Code CLI.
• The ucp CLI, for product discovery — set up below.
• Optionally the ucp skill, which makes catalog search more efficient.

1. Install

curl -fsSL -X POST https://delta-mandate-mcp.repyhlabs.dev/install | sh

Mints your identity and wires the MCP into Claude Code. Re-running the command above creates a fresh identity.

The minted identity is a bearer token: it authenticates you to this MCP and scopes the intents you can read back. It is not a signing key — this preview submits intents with Unsigned authorship, which the orchestrator accepts (see Authorship). Nothing here signs on your behalf.

Then run claude and type /mcp — you should see delta-mandate connected with 5 tools.

2. Set up Shopify UCP tools

You need the ucp CLI for product discovery. If it isn’t installed yet, Claude Code can set it up autonomously when needed.

The agent uses it to find products in Shopify’s global catalog via UCP and propose a variant — delta mandate extracts the evidence straight from that exact variant.

The ucp skill mentioned in the link above is optional; it makes the agent’s catalog search more efficient.

3. Use it

In Claude Code, type:

use delta-mandate to find me a black leather wallet under $50

What happens behind the scenes?

Two tools are in play — the delta-mandate MCP from step 1 and the Shopify ucp CLI from step 2:

1. Learn the policy language and find the product’s taxonomy category — get_language_guide + search_taxonomy (MCP).
2. Author and validate the policy — submit_policy (MCP), which returns a policy_id.
3. Search the Shopify catalog for a matching variant — the ucp CLI, which returns an identifier of the form gid://shopify/ProductVariant/42582963060798.
4. Submit the intent and then the proposal — submit_proposal with the policy_id and that identifier (MCP), which returns an intent_id. The intent is unsigned.
5. Read the verdict — get_outcome with the intent_id (MCP): success with the extracted evidence, or failure with the failed constraints.

Want to see a rejection?

Give an over-strict constraint (e.g. a price far too low). The agent proposes the closest match, the extracted evidence doesn’t satisfy the policy, and the policy engine returns failure — which is a valid result, not a malfunction: delta mandate caught a real mismatch between your policy and the product’s extracted evidence. You can also tell your agent to propose a different product or variant.

What it can (and can't) enforce

A policy can only constrain attributes the taxonomy exposes for a category, so some asks have no field — e.g. brand (“Nike”) isn’t a taxonomy attribute and can’t be enforced (the agent will tell you what it can enforce and what it has to drop). This preview works only against the Shopify catalog over UCP, using the Shopify Standard Product Taxonomy.