sil

███████╗██╗██╗
██╔════╝██║██║
███████╗██║██║
╚════██║██║██║
███████║██║███████╗
╚══════╝╚═╝╚══════╝

        Shop like you code

The shopping layer your agent runs for you.

Introducing Spec Driven Shopping: the way Agents should be shopping.

Website · Changelog · Security · Follow on X

Install

sil-openclaw is the OpenClaw plugin for sil. One install and your agent can shop the sil catalog on your behalf.

# ClawHub (recommended)
openclaw plugins install clawhub:@4gpts/sil

# npm
openclaw plugins install sil-openclaw

# Local checkout (dev / e2e)
openclaw plugins install /path/to/sil-openclaw

First run

Two lines and you're shopping:

1. Tell your agent:  "register me on sil"          (one browser sign-in, done)
2. Tell your agent:  "find me a mechanical keyboard under $100"

The agent loads the bundled skill on first commerce intent and takes it from there — searching, comparing, pulling up product detail, and handing back ready-to-buy links.

What sil is

A commerce catalog built for agents, not browsers. That's sil.

sil is a UCP (Universal Commerce Protocol) commerce service, and sil-openclaw wires it into your OpenClaw agent — so it can find products, compare prices and availability, pull up full detail, and surface a checkout link the moment you say buy. Powered by 4GPTs.

Today the plugin covers identity, catalog, and a personal multi-domain shopper. The rest of the UCP journey (cart, checkout, order, fulfillment) lands as those domains ship, and the same plugin grows with them.

See it in action

You, shopping through your agent — no store, no tabs:

you    find me a mechanical keyboard under $100, tactile switches

agent  searched the sil catalog — 3 strong matches:

         1. Keychron V3 (Brown)    $89   in stock
         2. NuPhy Air75 V2         $99   in stock
         3. Akko 3068B             $72   low stock

       want detail on any of these, or a checkout link?

you    tell me about the keychron

agent  Keychron V3 — QMK/VIA, hot-swappable, Brown tactile.
         options:   ABS / PBT keycaps · US / ISO layout
         featured:  PBT keycaps, US layout — $89, ships in 2 days
       → checkout:  sil.4gpts.com/c/kc-v3-pbt-us

you    perfect, send it

agent  done — here's your checkout link, ready to pay:
       sil.4gpts.com/c/kc-v3-pbt-us

You said three things. The agent did the shopping — searched, compared, fetched the detail, and handed you a ready-to-buy link. No browser, no forms, no fourteen open tabs.

Turn your OpenClaw into your personal shopping expert

So far your general agent did the shopping. Go one step further: create one persistent shopper that learns you once, then shops any niche like a specialist — it researches how to buy well in each niche the first time you shop it, and gets sharper every session.

Spec-Driven Shopping (SDS). Say "set up my shopper" and the plugin runs a light interview (≤10 questions) to shape how it should behave (the persona) and capture your cross-niche facts and the rules it must never break — and that's the whole setup. No niche is researched up front. From then on you just shop. The first time you ask for something in a new niche it classifies what you're buying, researches that niche deeply on the spot (the full how-to-buy-well: fit, gearing, geometry, the trade-offs — announced, so you can correct it), and derives the dimensions every request in that niche is decomposed along — minting a reusable domain. Ask for that niche again and it reuses what it learned; ask for a new one and it learns that too, keeping you the same across all of them. On every query it refreshes the domain knowledge from the web, decomposes what you asked, and learns your facts and taste as it goes (your measurements, your budget, a brand you won't touch) — captured once, never re-asked. It recommends with a "why" that cites what you asked, a fact it remembered, and a niche mechanic you'd never have named. Nothing is created until you say so, and a second shopper is refused — there is exactly one of you. Creating and running it is local and offline, no sign-in required.

1 — One light conversation creates your shopper (no niche research yet):

you    set up my shopper

agent  Love it — let's shape it lightly. Should it shop patient and
       thorough, or fast and decisive? Anything it must always do,
       or never do?

you    thorough, and never recommend anything I can't return

agent  Got it — I'll keep your cross-niche rules (free returns only)
       and learn each niche the first time you shop it. Nothing to
       research up front.

         Your Shopper — a thorough, returns-first buyer
         · learns any niche on first shop · keeps your facts across all of them

       shall I create it?

you    yes, create it

agent  done — your shopper is live. Just tell it what you want;
       it learns the niche on the fly.

2 — It learns a niche the first time you shop it, and keeps you across all of them:

you    find me a road bike, ~€1500

agent  New niche for us — road cycling. Let me learn how to buy one
       well (fit from height AND inseam, endurance-vs-race geometry,
       gearing, the comfort/speed trade-offs) — and two numbers I'll
       keep so I never re-ask: your height and inseam.

you    178cm, 81cm inseam. Secondhand is fine

agent  searched sil — endurance, 54cm for your fit, ~€1500:

         1. Alloy endurance bike, 54cm, 105 (used)   €1420
         2. Carbon endurance bike, 54cm, 105 (used)  €1610

       Why #1: for your 178/81 fit a 54cm endurance frame sits where a
       first buyer wants it — and it honours your "free returns only"
       rule. Want detail or a link?

you    later — find me a backpacking tent for two, under €250

agent  Different niche — camping. Learning how to buy a backpacking
       tent (season rating, packed weight, floor area)… and your
       returns rule carries straight over, no need to re-ask.

You teach it once; it stays sharp. It keeps your facts and taste, web-refreshes its niche knowledge each visit, and never re-asks what it already knows. See what your shopper knows and which domains it has learned, look at how a domain is set up, forget one you're done with — or refine the shopper or a domain from what it watched you reject, so it gets sharper every session.

What it remembers, and where it lives. Your shopper learns your shopping-relevant facts as it goes — measurements, budget band, the rules it must never break, brand likes/dislikes — captured only when a request actually needs them (never a big up-front form) so it doesn't re-ask. All of it is stored locally on your machine ($SIL_DATA_DIR/shopper/, owner-only 0600), per-user (one shopper, many domains). It is never pooled across users and never sent to a server for training or aggregation — creating and running your shopper is local and offline. The niche web-refresh reads public sources to keep the domain knowledge current; it does not upload anything about you. Inspect exactly what your shopper holds with sil_profile_get, and forget one domain it learned with sil_profile_remove.

Tools

Namespaced sil_* so they never collide with other plugins. Your agent calls them for you — you just say what you want.

Identity

Catalog

Your shopper — create one shopper, then manage the domains it learns

Skills

The plugin ships one bundled skill — sil-shopping 🛒 — that your agent loads automatically the first time you express a shopping intent. You don't invoke it; it's the playbook that makes the tools work well together:

• Routes intent to the right tool. "find me a keyboard" → sil_search, "look these up" → sil_product_get, "who am I?" → sil_whoami, "sign me up" → sil_register, "set up my shopper" → the create flow, "what does my shopper know / forget the grocery domain" → the sil_profile_* tools.
• One shopper, many niches — minted on the fly. Asked to set up shopping, it runs a short two-touchpoint onboarding (how your shopper should behave + your cross-niche facts and hard rules) and creates one shopper — creating nothing until you explicitly endorse the draft. From then on it shops any niche: it classifies what you're buying, reuses a niche it has already learned or researches a new one on the spot (announced, so you can correct it), and derives how to decompose every request — learning your facts and taste as it goes.
• Recovers the right way. Every tool reports a status; the skill follows that tool's own recovery hint — re-register, fix the query, or retry — instead of guessing a fix that won't work.
• Keeps prices honest. It treats price, availability, and checkout links as point-in-time and re-checks an item right before you buy, so the link you get is the link you pay.

Because the skill ships inside the plugin, installing the plugin installs the skill — there's nothing extra to set up.

Security

The plugin holds your sil credentials and transacts on your behalf. The full disclosure — what it touches, what it stores, and how to report an issue — is in SECURITY.md.

Developing

pnpm install
pnpm build       # pnpm clean && tsc → dist/
pnpm test        # vitest (unit + integration)
pnpm typecheck   # tsc --noEmit

Releasing is two steps: pnpm version <patch|minor|major> (bump → sync manifest → cut changelog → test → tag → push), then pnpm release (build → pack → npm sil-openclaw + ClawHub @4gpts/sil — the same contents, re-packed under each registry's name). Release notes in CHANGELOG.md. Adding a tool is three steps, enforced by a drift-guard test.

Built by 4GPTs · Apache-2.0 · @4gpts on X