# LifeHack 2026 > Woven — an MCP commerce app for ChatGPT and Codex that builds compatible, in-stock carts and requires explicit confirmation before simulated Visa authorization. - **Type:** MCP server - **Install:** `agentstack add mcp-ducksss-lifehack-2026` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [Ducksss](https://agentstack.voostack.com/s/ducksss) - **Installs:** 0 - **Category:** [Integrations](https://agentstack.voostack.com/c/integrations) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [Ducksss](https://github.com/Ducksss) - **Source:** https://github.com/Ducksss/LifeHack-2026 - **Website:** https://visa-woven.vercel.app/ ## Install ```sh agentstack add mcp-ducksss-lifehack-2026 ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About Woven Everything works together. An agentic-commerce app inside ChatGPT and Codex, backed by bounded mission orchestration and deterministic checkout controls. One request becomes a complete cart; only the user can approve the exact purchase. Ask once. Review once. Confirm once. Open Woven · Run the live demo · WebMCP challenge kit · Install Woven · Run locally · Product requirements · Technical architecture · Devpost submission kit · 3-minute pitch · Live demo · Report an issue > [!IMPORTANT] > Woven is a hackathon prototype. Merchants, inventory, prices, and Visa > authorization responses are seeded or simulated. It collects no card > credentials and cannot make a live charge. Table of contents About the project Why ChatGPT and why a plugin? How it works The three-minute story Product gallery Architecture Built with Getting started Usage Trust and safety Testing Roadmap Contributing License Contact Acknowledgments ## About the project **Search gives you links. Buying still takes work.** Imagine taking two first-time campers away for a rainy weekend. You still have to size the shelter, duplicate the sleep gear, check weather ratings, keep the packed kit inside one car boot, find one store with everything in stock, and rebuild the cart at checkout. Woven finishes that job inside the AI workflow the user already uses. Ask ChatGPT or Codex once and Woven returns complete carts from one pickup location, opens a Choice Center to compare five options, shows why every item works together, rechecks the price and stock, and waits for a separate human confirmation before authorization. The in-chat MCP App and browser-native WebMCP tools are interaction layers, not the whole backend. Woven's Node.js service owns mission routing, a bounded orchestration workflow for non-camping requests, deterministic cart verification, SQLite state, and the checkout trust boundary. The canonical mission is deliberately concrete: > I need a complete rainy-weekend camping kit for 2 first-time campers. Keep it > under S$300, fit it in one car boot, and make it pickup-ready today. ### Why it is different | Product search | Woven | | --- | --- | | Ranks individual items | Ranks complete, compatible carts | | Leaves stock and pickup implicit | Uses current demo stock and pickup timing | | May mix merchants | Returns one pickup location per cart | | Makes the user reconstruct compatibility | Shows a proof for every component | | Forces a single recommendation | Compares five complete choices and lets the user rerank them | | Blurs recommendation and purchase | Requires an exact, expiring confirmation | (back to top) ## Why ChatGPT? Why a plugin? The shopping request already exists in the conversation. A standalone app would make the user repeat it, while Woven can turn that same request into bounded merchant actions and return a structured checkout for review. The primary product is a real MCP App/plugin inside ChatGPT or Codex. The model can ask Woven to build carts, but app-only tools and a direct user click protect checkout. The `/demo` browser route is a stage-safe rehearsal of that experience: a clearly labeled simulated chat host that types the canonical request, shows every MCP tool call live, and drives the same backend. It is marked “Simulated” on screen and is not a second product. The `/webmcp` route adds seven imperative, top-level WebMCP site tools to the same page. An agent can start or inspect a mission, open the shared comparison UI, select a cart, apply a merchant-approved swap, refresh current offers, and verify a receipt. Identity, checkout preview, and purchase confirmation are deliberately absent: those remain direct human actions in the visible page. MCP supplies the host connection, tool contract, private widget metadata, and embedded UI. After `start_mission` reaches Woven, the server—not the host model—routes the request, runs any required orchestration, classifies evidence, and decides whether a cart is checkout-eligible. (back to top) ## How it works 1. **Ask once.** An MCP host or the WebMCP-enabled site calls `start_mission` with the user's mission and any structured constraints. 2. **Route and orchestrate.** The canonical camping request uses the deterministic engine. Other categories enter a fixed LangGraph.js workflow that interprets a `MissionSpec`, discovers connected offers and cited research, normalizes, composes, verifies, and retries at most once. 3. **Build complete carts.** Deterministic server rules reject incompatible, unverifiable, or unavailable offers, stay under budget, and return up to five one-location options. Web findings remain research-only. 4. **Compare and choose.** The Choice Center reranks by value, pickup speed, rain protection, or preferred area; optional saved preferences stay local. 5. **Show the proof.** The MCP App widget explains tent capacity and rain rating, two sleeping bags, two mats, lantern protection, first-aid coverage, packed volume, live demo stock, total, pickup plan, and approved substitutions. 6. **Verify the demo identity.** A provider-style page returns a single-use code; Woven validates state + PKCE and creates a short-lived server session. 7. **Review once.** The server rechecks price and stock and creates a ten-minute mandate bound to the identity session, merchant, cart version, and amount. 8. **Confirm once.** A private nonce, mandate hash, and idempotency key are verified before the simulated authorization, merchant order, and signed server-verifiable receipt. The AI recommends. The user chooses. Woven binds the exact terms. ### Bounded orchestration backend The canonical camping demo above is still deterministic and unchanged. Woven also includes a credential-dormant TypeScript orchestration POC for non-camping missions. This is a backend layer behind the same `start_mission` tool, not a second MCP server or a host-model prompt convention: - a fixed LangGraph.js flow interprets a validated `MissionSpec`, discovers connected and web offers in parallel, normalizes, composes, verifies, retries at most once, ranks, and persists; - connected catalog offers can form checkout carts only after the server verifies every hard requirement, compatibility link, quantity, stock, budget, merchant, and pickup location; - web results remain cited research leads with checkout disabled; and - preview and confirmation rebuild connected carts from current SQLite catalog rows, preserving the existing identity, nonce, mandate, idempotency, and atomic inventory protections. The POC uses the OpenAI Responses API with schema-constrained output, `gpt-5.6-terra`, medium reasoning, built-in web search, and `store: false`. No API credential is required for the camping demo or CI. Without `OPENAI_API_KEY`, a non-camping mission returns retryable `AGENT_UNAVAILABLE` rather than fabricated results. Developers who already have a key can run the opt-in live matrix with `npm run eval:agent`. (back to top) ## The three-minute story ```text Ask once ↓ Receive complete one-merchant carts ↓ Compare five choices and tune priorities ↓ See why everything works together ↓ Verify the simulated demo identity ↓ Review the exact merchant, items, pickup, and total ↓ Confirm once ↓ Receive a simulated Visa result and pickup receipt ``` The working prototype includes a **simulated connector-style identity check** before checkout. It is server-enforced but is not a Visa login, KYC, or a real identity-provider integration. It remains separate from final purchase confirmation and never collects Visa credentials or card data. See [`script.md`](script.md) for the exact narration, stage actions, fallback path, and the working identity handoff. ### Render the judge video The repository includes a three-minute Remotion composition that follows the authoritative script, uses the verified Woven product frames, and labels both identity and Visa authorization as simulated. ```bash npm run video:studio # preview and scrub the composition npm run video:render # output/Woven-Judge-Video.mp4 ``` The sidecar captions are in `video/Woven-Judge-Video.srt`. The bundled voice is AI-generated with ElevenLabs and can be replaced by swapping the files in `public/woven-video/voiceover/` without changing the timeline. (back to top) ## Product gallery Born inside the chat.One message becomes a live MCP app — visible tool calls, then five complete carts in the Choice Center. A visible authorization boundary.The exact merchant, pickup, items, and total are bound before confirmation. A separate identity gate.The provider-style simulator creates a short-lived server session without collecting Visa, card, or payment credentials. A pickup-ready result.The user receives a simulated authorization result and receipt. A controllable live demo.Inventory, pricing, declines, order failures, and the audit trail are visible on stage. (back to top) ## Architecture Woven is one deployable Node.js service with five explicit layers: MCP/HTTP interaction, mission routing and bounded orchestration, deterministic commerce verification, SQLite checkout state, and the payment adapter. The browser fallback and merchant desk enter the same backend as the MCP App. The real payment integration boundary is isolated in `src/payment.ts`; the current adapter intentionally fails closed unless `PAYMENT_MODE=simulated`. LangGraph.js coordinates the non-camping workflow; OpenAI supplies structured interpretation and cited web research; deterministic TypeScript rules decide whether connected offers satisfy the mission; and the SQLite store revalidates them before checkout. MCP carries inputs and results but does not own those backend decisions. Detailed tool contracts, state transitions, cart rules, and trust boundaries are documented in [the architecture guide](docs/architecture.md). The live [target-architecture explainer](https://visa-woven.vercel.app/architecture) visualizes those boundaries without implying a current Visa integration. (back to top) ## Built with - [TypeScript](https://www.typescriptlang.org/) and [Node.js](https://nodejs.org/) - [React](https://react.dev/) and [Vite](https://vite.dev/) - [Tailwind CSS](https://tailwindcss.com/) and [shadcn/ui](https://ui.shadcn.com/) - [Model Context Protocol SDK](https://modelcontextprotocol.io/) - [MCP Apps](https://github.com/modelcontextprotocol/ext-apps) - [Express](https://expressjs.com/) - Node's built-in [SQLite](https://nodejs.org/api/sqlite.html) - [Zod](https://zod.dev/) - [LangGraph.js](https://docs.langchain.com/oss/javascript/langgraph/) for the bounded open-world workflow - [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses) for optional structured interpretation and cited web research - [Remotion](https://www.remotion.dev/) for the three-minute judge video (back to top) ## Getting started ### Prerequisites - Node.js 22.5 or newer - npm ### Installation For the exact Codex desktop, Codex CLI, and ChatGPT Developer Mode steps—with expected output and troubleshooting—use the **[installation and verification guide](docs/INSTALLATION.md)**. ```bash git clone https://github.com/Ducksss/LifeHack-2026.git cd LifeHack-2026 npm ci npm run check npm start ``` Open the local surfaces: | Surface | URL | | --- | --- | | Landing page | | | Buyer fallback | | | WebMCP workspace | | | Demo identity connector | | | Merchant desk | | | Install guide | | | MCP endpoint | | | Health check | | The verified public deployment is available at [visa-woven.vercel.app](https://visa-woven.vercel.app), with the buyer demo at [/demo](https://visa-woven.vercel.app/demo) and merchant desk at [/merchant](https://visa-woven.vercel.app/merchant). Its SQLite database lives in Vercel's temporary filesystem, so demo state may reset after a cold start or redeployment. Add `?loop=true` to `/demo` for an unattended visual walkthrough. It compares and reranks carts, returns to the selected kit, shows the compatibility proof and approved swap, completes the server-enforced simulated identity handoff, shows the verified state and host reply, then fades back to the start. It never creates a checkout mandate or confirms a purchase. Open `/webmcp` in a current WebMCP-capable ChatGPT or Codex browser session to discover the seven page tools. The readiness card and shared activity rail make agent actions visible; the browser and MCP transports enter the same server-owned mission router. ### Connect the ChatGPT app Follow the [ChatGPT connection guide](docs/INSTALLATION.md#3-connect-woven-to-chatgpt). It covers the required public HTTPS endpoint, Developer Mode setup, discovered tools, and expected widget result. ChatGPT cannot connect to localhost. Current OpenAI references: [plugin quickstart][openai-quickstart], [MCP server guide][openai-mcp], [ChatGPT UI guide][openai-ui], and [connection guide][openai-connect]. ### Connect the Codex plugin Follow the [local Codex installation guide](docs/INSTALLATION.md#2-recommended-install-in-codex-desktop). The repository includes a **Woven Local** marketplace entry; its bundled stdio server returns a self-contained Choice Center; port `8788` remains available only for the simulated identity handoff. (back to top) ## Usage ### Happy-path demo Use the complete [three-minute stage script](script.md). The working happy path is: 1. Open `/demo` — the simulated chat host types the canonical request, plays the `start_mission` activity, and opens the Choice Center with five ranked carts. 2. Compare carts, tune a priority or pickup area, then select a kit and inspect its compatibility proof and pickup plan. 3. Click **Review checkout**, then **Verify demo identity**. 4. On `/identity`, click **Continue as Chai**; return and check the result. If authorization is still pending, the same status action reopens the handoff. 5. Review the exact, expiring mandate and click **Confirm S$231.00**. 6. Show the simulated Visa result and signed receipt. ### Trust/failure demo Use `/merchant` to select a scenario, then replay checkout: - **Stockout** or **Price change** invalidates an old preview. - **Auth decline** creates no merchant order. - **Order failure** enters reversal after simulated authorization. - **Reset demo data** restores stock and clears the local run. ### Catalog updates The merchant desk exports and imports a deliberately small CSV contract: ```csv offer_id,price_sgd,stock trailhaus-funan-th-storm2,89.00,4 ``` Imports update only known offers and reject malformed prices, negative values, and fractional stock. A demo upload cannot silently create incompatible SKUs. ### Commands ```bash npm run dev # Vite UI development server npm run dev:server # API/MCP server with restart-on-change npm run build # production bundle + TypeScript validation npm test # domain, security, idempotency, failure, and CSV checks npm run check # full test and build gate npm run mcp # stdio MCP transport used by the Codex plugin npm run video:studio # preview the three-minute judge video npm run video:render # render output/Woven-Judge-Video.mp4 ``` ### Configuration | Variable | Default | Purpose | | --- | --- | --- | | `PORT` | `8787` | HTTP server port | | `BASE_URL` | `http://localhost:8787` | Public widget asset/CSP origin | | `WOVEN_DB` | `./data/woven.db` | SQLite demo state | | `PAYMENT_MODE` | `simulated` | Guardrail; every other value fails closed | | `OPENAI_API_KEY` | unset | Optional; enables non-camping orchestration. Camping, build, and CI remain credential-fr … ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [Ducksss](https://github.com/Ducksss) - **Source:** [Ducksss/LifeHack-2026](https://github.com/Ducksss/LifeHack-2026) - **License:** MIT - **Homepage:** https://visa-woven.vercel.app/ Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-ducksss-lifehack-2026 - Seller: https://agentstack.voostack.com/s/ducksss - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.