They wait you out. Outwait waits longer.
An agent that chases a company for money they owe you — for weeks, on its own — and taps you only for the ten seconds that genuinely need a human.
Live app · Watch the demo · Check it yourself · Build log · The three boundaries · How it works · Running it
Nobody refuses you. They just make getting your money back take longer than you are willing to spend on it.
In England and Wales alone, 4,706,470 tenancy deposits are held in protection schemes, worth a record £5.53 billion. The average deposit is £1,175 — real money to the person who paid it.
In 2024/25 those schemes carried out 46,950 adjudications. That is 1.00% of protected deposits.
Ninety-nine percent of deposits never reach a dispute.
That is not because ninety-nine percent of landlords return the money cleanly. The same briefing lists what disputes are actually about — cleaning (54%), damage (49%), redecoration (31%) — the ordinary end-of-tenancy arguments that happen everywhere, not rare misconduct. And when a case is adjudicated, the landlord is rarely handed the lot: industry analysis of the TDS annual report puts a 100% award to the landlord at around 5% of insured cases, with most ending in a split.
So the money is not usually lost because the tenant was wrong. It is lost because pursuing it is a part-time job, and a normal person quits somewhere around the third email.
Sources: TDS Statistical Briefing 2025 (deposit counts, value, adjudication rate, dispute reasons — figures for England and Wales, year to March 2025). Award-split figure reported from the TDS annual report by The Accommodation Bureau. Deposits are the clearest public dataset; the same shape applies to refunds, delayed payouts and wrongly charged bills, where no scheme keeps score at all.
A claim is not one hard task. It is a long chain of small ones, and every link is a place to give up:
| The step they add | What it costs you |
|---|---|
| Fill in the online form | Ten minutes, and a reference number |
| Wait ten working days | Two weeks of remembering to check |
| Send the same document again | The suspicion that nobody read the first one |
| "Please call us between 10 and 4" | A working day you do not have |
| Escalate in writing to another team | Starting over, with a new person |
None of it is difficult. All of it is long. You do not have to outsmart them. You have to outlast them — and a machine does not get bored.
You forward one email. That is the entire setup — no form, no account required to start.
The forward hands over everything a claim needs and a form would have made you retype: the company's real reply-to address, the reference number, the dates, the amounts and the history. From there it opens a case, researches the company's own published policy, writes to them, waits a week, writes again, registers on whatever portal is required, and escalates — for as long as it takes.
The board answers two questions and refuses to ask you anything else: is anything happening, and does it need me?
Open one and you get the whole story as a timeline in plain English — no status codes, no jargon, no log to decode.
The claim this product makes is checkable, so the landing page lets you check it. Name a company you have actually dealt with and the same code that runs on a real claim goes and reads their own pages, finds the deadline they set for themselves, and shows you the sentence it came from with a link. If they publish nothing, it says that instead.
No model is in that path. The number is found by anchored regex over the page text and checked against fixtures of real policy wording — and, more importantly, against the sentences that must not count. Ten kinds of sentence look exactly like a promise and are not one, and every one of these is real text pulled off a real page:
| What it looks like | What it actually is |
|---|---|
| "you must return the item within 14 days" | a deadline on you |
| "you can return new and unopened products within 365 days" | your return window, offered not obliged — IKEA |
| "within 30 days of the purchase date" | a return window again — H&M |
| "it can take up to five working days for your bank to process it" | your bank's clock — Argos |
| "PayPal refunds can take up to 30 days" | a payment provider's window — Argos |
| "Mobile phone billing — up to 60 days for the statement" | the carrier's clock — Apple |
| "Card approval cancellation within 3-5 business days" | an authorisation hold releasing — Airbnb |
| "a refund or flight voucher ... to be used within six months" | a voucher's shelf life — easyJet |
| "within 10 days prior to your flight" | an eligibility window running backwards — Ryanair |
| "@Carla1984 Automatic refunds can take 6 weeks" | a community forum post — Sky |
All ten are refused. Each is a fixture: 41 cases in
backend/experiments/promise-extraction.test.mjs, runnable with node and no setup.
On a real claim that number lands on the case with the date it runs out, and that sentence is what gets quoted back at them.
The page keeps a public ledger of everything it has been asked to read. It is the product's own research, not a dataset we loaded: every row is a live page, read by the same code, linked so the reading can be checked. It is a live Convex query, so a company you look up joins the list in front of you without a refresh.
20 of 28 set themselves a deadline. They wrote it down. Only eight — Sky, Vodafone, Uber, Deliveroo, Just Eat, British Airways, Booking.com and easyJet — name no number at all on their own refund pages. Of the twenty that do, the middle promise is 14 days, and the range runs from Next's one working day to Airbnb's fifteen.
That is the thesis with a number on it, and it is not the one we expected. The deadline usually exists. It is just four clicks into a terms page, on the day you are too tired to look, and nobody is counting from it — so "we are still processing it" holds, indefinitely, against a promise the company made itself. For those twenty the product goes and finds that sentence and counts the days from it. For the other eight there is nothing to count from: a company that never names a date can never be late.
Method: their own site only, the pages a search for their refund policy actually returns — the same place a customer would look. Every row here has been read at least twice, because one read was not enough: a first pass reported eight of these companies as publishing nothing and a second pass found a published deadline on every one of them — Currys, Apple, Samsung, Wayfair, EE, H&M, HSBC and Nationwide. Search does not return the same pages twice, and the miss was ours, not theirs. "Nothing published" now means nothing found on two separate reads, and every reading is taken again once it is a week old. Community and forum URLs are excluded, because a post by another customer is not a commitment.
Every agent has a human. The only real question is where that human stands — and this product answers it three times. Each answer is a line in the code, not a promise in a prompt.
A complaints portal wants an account. The agent registers in its own name, with its own email address, then reads its own verification code out of its own inbox and carries on.
It never holds a credential of yours, because it never needs one.
Anything binding — a formal complaint, a number, a settlement — is written and then held as a draft. It goes out when a person replies "yes" to an email, and there is no code path that sends it without one. The guard is structural, not an instruction the model is asked to respect.
When a step needs your login, it does not ask for your password.
It sends one link. You tap it on your phone, a real browser opens inside the page, you type your password into the site itself, and the agent carries on in that same session. The session is created the moment you tap — not when the email was sent — because a browser left open waiting for you would be dead by the time you arrived.
That is a live browser, in the page, with the clock running. Your password is typed into the real site; on every other screen watching that session it renders as dots. Afterwards the login is saved against that company, so it never has to ask you twice.
Ten seconds of your attention, once per company, forever.
forward an email
│
▼
signed webhook ──► verify ──► store ──► 204 (answer fast, think after)
│
▼
read the letter ──► open a case + its tracks (one file calls a model)
│
├──► research the company's own policy Firecrawl
├──► write to them, wait a week, write again durable workflow
├──► register on a portal as itself Firecrawl + own inbox
└──► when it truly needs you ──► one email AgentMail
│
▼
reply "yes" · send a photo · tap a link
Everything the demo depends on is deterministic. One file calls a model
(convex/agent/model.ts, two named jobs) and it only ever turns prose into structured fields.
The case shape, which tracks open, what the board says and what gets sent are all plain
mutations — so the same email always produces the same case.
Not a database with a server bolted on. The parts of this that would normally be infrastructure are Convex features, which is why a solo build could reach this scope.
| What | Where |
|---|---|
Durable workflows — step.sleep for 7 days, three times, then escalation. A chase that survives restarts and costs nothing while it waits. |
convex/workflows/chase.ts · @convex-dev/workflow |
| Realtime queries — one subscription drives the whole board. Mail lands, the screen moves. Nobody refreshes. | convex/cases.ts |
| HTTP actions — the inbound mail webhook, with hand-written Svix signature verification (Web Crypto, because HTTP actions are not Node). | convex/http.ts · convex/lib/svix.ts |
| Scheduled functions — answer the webhook in 204, then think. A slow webhook gets retried, and a retry would open the case twice. | convex/http.ts |
| Auth — sign in with a six-digit code, no password anywhere. | convex/auth.ts · @convex-dev/auth |
| Presence — three flatmates, one deposit: see who else is on a case and who is holding the wheel. | convex/presence.ts · @convex-dev/presence |
| File storage — reply to an email with a photo of a receipt and it becomes evidence on the case. | convex/mail/client.ts |
| Static hosting — the whole frontend served from the same deployment. | @convex-dev/static-hosting |
Three registered components, plus auth. convex/lib/status.ts derives case status from the
rows themselves rather than letting each blocker assign it — which removed a whole class of
bug where clearing one blocker wrongly cleared the case.
Two different jobs, and the second one is the interesting one.
- Reading.
/v2/searchfinds the company's own published pages and/v2/scraperenders and reads them, so a letter can quote their own deadline back at them. Guarded twice: text is only attached when the page hostname matches the company on the case — after an unrelated business's refund policy was once stored as if it were theirs — and the number itself comes out by regex, never a model. This is also what powers Check it yourself, where anyone can point the same code at any company and watch it work, cached by hostname so it costs a credit once. - Operating.
/interactdrives a real browser to fill in portals — andinteractiveLiveViewUrlis what makes boundary 3 possible: the same live session handed to a human, mid-flight, then handed back. Named profiles persist the login so it is asked for once per company.
A finding worth passing on: the profile write is asynchronous. Reading a profile back too soon returns an empty one — 0 of 3 sessions reused the login with no wait, 3 of 3 with a settle. That only showed up by measuring it.
The product reaches people by email, so email is not a notification channel here. It is the interface.
- Its own inbox and address — which is what lets the agent sign up for things as itself, and read its own verification codes.
- Inbound webhook, Svix-signed — a forwarded email becomes a case. Unsigned and forged deliveries are rejected.
- Threads as case files — a reply lands back on the right case with no reference number to quote.
- Drafts as the approval gate — boundary 2 is an AgentMail draft that is held until a human says yes.
- Sign-in codes — the same inbox that chases companies also sends your login code.
Verification codes are pulled out of mail by anchored regex rather than a model, checked against a fixture set of real and decoy messages, because sign-in should not depend on a model's mood.
openai/gpt-5.6-luna via OpenRouter, in exactly one file, doing two named jobs: turn a
forwarded email into structured case fields, and classify a reply as yes / no / neither.
Everything downstream is deterministic on purpose. A demo that depends on a model agreeing with you is a demo that fails live.
Three independent npm projects, no workspaces.
backend/ the Convex app — schema, functions, the agent loop, workflows
frontend/ Vite + React + shadcn/ui, builds to frontend/dist
component/ reserved for a publishable Firecrawl /interact component
# backend
cd backend
npm install
cp .env.example .env.local # then fill it in
npx convex dev
# frontend, second terminal
cd frontend
npm install
npm run devDeploy backend and static site together, from backend/:
npm run deployCredentials live only as Convex environment variables, never in the repo:
FIRECRAWL_API_KEY, AGENTMAIL_API_KEY, AGENTMAIL_WEBHOOK_SECRET, OPENROUTER_API_KEY,
MODEL_ID. Names are in backend/.env.example.
Honest about today, and each of these is the natural next build rather than a hole.
Phone calls. When a company insists on a call, it puts the same thing in writing instead — usually the stronger move, since writing leaves a record. A voice agent that sits on hold and reads out the reference number is the obvious extension, and hold time is exactly the kind of cost a machine should absorb.
A visible schedule. The chase already runs on a durable workflow that sleeps for weeks. Surfacing that as a dated plan — "I write again on the 26th; if nothing by the 3rd, I file with the ombudsman" — turns waiting from an absence into something you can see coming.
Watching the promise change. A company quietly editing its own turnaround from 14 days to 30 is exactly the kind of thing nobody notices, because the old page is gone the moment it is saved. The sentence and the date it was read are already stored, so keeping every version and telling people when one moves is a small step from here — and it would make a category of quiet rule-changing visible for the first time.
Sessions that outlive the site's own. A saved login lasts only as long as the company's session does — minutes at a bank, months at a small portal. Re-authenticating without asking again, where the site permits it, is the difference between ten seconds once and ten seconds occasionally.
Companies with good self-service. Amazon already refunds you in two clicks and needs none of this. The product is aimed squarely at the ones that do not, and the addressable gap is every company that made a process long on purpose.
This is not legal advice and never claims to be. The claim is narrower and, we think, more useful: it chases, it keeps a record, and it does not get bored.
Built for the Convex All Gas Hackathon.
Convex · OpenAI · Firecrawl · AgentMail