Stow | Devpost

Inspiration

The OpenAI WebMCP showcase is full of tools that hand a problem to an agent and let the human watch. We wanted the opposite. Packing a real bag is an asymmetric co-op: the human brings what cannot be left, the agent brings what the
threshold will not allow. Neither side should win unilaterally. An agent that unpacks a violin is wrong. A human who ignores a TSA rule is also wrong. We built Stow to make that negotiation visible on one shared page.

What it does

Stow is a packing cockpit for a human and their agent. They share one bag, one scale, and one manifest log.

The human pins items that cannot be lost (a violin, a toddler blanket, a photo album).
The agent proposes a repack that clears airline, TSA, and go-bag rules.
propose_repack is constitutionally unable to drop a pin. That contract is the product.
The human accepts or rejects the proposal. The agent does not touch the bag without consent.
seal_bag is irreversible and only succeeds when zero blocking violations remain.

Three presets cover the failure modes: a Ryanair weekend (10 kg cabin, illegal lotion, too much mass), a family cabin (toddler + liquids + 10 kg), and a wildfire go-bag (72 hours, no documents = block).

How we built it

WebMCP, native first. document.modelContext.registerTool runs in ChatGPT's in-app browser and in Chrome 149+ with chrome://flags/#enable-webmcp-testing. Chrome Labs' polyfill loads only as a fallback so judges without the flag still see the same tools.
Progressive tool registration. Tools register and unregister via AbortSignal as bag state changes. accept_proposal only exists while a proposal is pending. add_to_bag disappears after seal_bag.
Declarative + imperative API. A set_trip form uses toolname / tooldescription / toolautosubmit so a human or an agent can fill it. All other tools are imperative.
Actionable errors. Every tool returns a string the agent can recover from: Unknown item, Cannot remove "X": it is pinned, Cannot seal: cabin bag is 12.4 kg over a 10 kg limit, No pending proposal.
Veto-aware proposer. A deterministic greedy algorithm drops unpinned items by grams-per-utility until blocking rules clear. Pinned mass alone over the limit is reported in plain language instead of silently failing.
One source of truth. Human clicks and agent tool calls both go through the same Store. The UI re-renders from the same state the manifest log reads.
Stack: TypeScript, Vite, Vitest. No React, no Tailwind, no backend. Apache-2.0.

Challenges we ran into

License detection. The first LICENSE draft was a paraphrased Apache-2.0; GitHub's licensee flagged it as Other/NOASSERTION. Fixed by pulling the canonical apache.org text and adding a NOTICE file.
GitHub Pages base path. The polyfill script tag was hard-coded to /webmcp-polyfill.js, which 404s under a project subpath. Switched to new URL('webmcp-polyfill.js', document.baseURI).href so it resolves relative to the page.
Demo handoff visibility. The first cut put the agent's proposal card in the right-hand column, below the fold. Judges would have missed it. Lifted it to a brass-bordered banner above the bag with the line "Agent proposal — pins are
kept" so the human/agent handoff is the first thing the eye lands on.
Video height for x264. Headless Chrome captured 1440x813; x264 requires even dimensions. Used scale=1440:trunc(ih/2)*2 instead of cropping.
Tool registration race. Loading the polyfill then immediately calling registerTool could collide with the in-flight script. Solved by serializing boot (await loadScript then syncTools).

Accomplishments that we're proud of

A WebMCP app you can hand to a non-technical person. Load a preset, pin an item, ask the agent to repack. No MCP, no ChatGPT account, no flag required. The on-page Agent Tools console lets a judge exercise every tool from Chrome
DevTools.
A veto the agent cannot override. propose_repack is the only repack tool. It is non-mutating and refuses to drop a pin. The companion test (propose.test.ts) asserts the violin survives any number of over-weight scenarios.
Actionable, recoverable errors. Every tool returns a string a future agent can act on. No Error: undefined. The wrapper tests pin those strings.
23 passing tests, 4 files, one engine. Bag invariants, airline/TSA/go-bag rules, veto-aware proposer, and progressive tool registration. The CI runs them on every push.
Apache-2.0. Canonical license text, NOTICE file, public repo, public Pages, public video. The "About" sidebar shows Apache-2.0.

What we learned

WebMCP's biggest win is shared state, not tool calling. Once human and agent write to the same store, the UI stops being a spectator sport.
Annotations matter. readOnlyHint and destructiveHint change how an agent decides between a propose_repack and a seal_bag. The Chrome DevTools WebMCP panel surfaces them; ignoring them makes the agent overconfident.
The polyfill is a real product surface. Even on Chrome 150 the flag isn't always on, and judges may not have ChatGPT's in-app browser open. Falling back to Chrome Labs' polyfill is the difference between a one-tool demo and a
working app.
Progressive registration is the right pattern. Aborting and re-registering tools as state changes is more honest than letting accept_proposal silently return No pending proposal for every call. The agent can see what it is allowed
to do.
The Devpost rules want a recognisable snippet. search_products with the exact schema from the rules is in the repo. Subtle but the judges' checklist would have flagged a missing one.

What's next for Stow

Live airline data. Pull the real Ryanair 10 kg / BA 23 kg limits from a feed instead of the static table.
Multi-human packers. Two people packing a shared family bag, each with their own pins.
A "trust meter" the agent earns. Track which proposals the human has accepted and start suggesting riskier repacks for items the human keeps un-pinning.
Real boarding-pass constraints. Seat-pitcher rules, gate-check bag size, lithium battery watt-hour caps.
A watch-your-language layer for agents. Detect when a tool result includes user-influenced content and warn the agent before it mutates the bag.
A backend for multi-device. Right now the bag lives in the page. With backend MCP the human and agent could keep packing on the train while the laptop sleeps.