QuickBooks sends reminders. Cashflow Agent remembers the conversation.
An AI agent that reads your customer threads and writes the follow-up that actually gets you paid.
Built for the Hermes Agent Accelerated Business Hackathon — Nous Research × NVIDIA × Stripe
Small business owners waste hours chasing payments. QuickBooks fires off the same generic reminder to everyone. Customers ignore them.
But a follow-up that says "Hey John, you said you'd pay Friday — I still haven't seen it" actually gets a response. The problem is nobody has time to write those by hand.
We analyzed 15,000+ small business pain signals across Reddit, X, and forums. Getting paid is the #1 pain with the fewest solutions — the market is saturated with lead-gen tools, but barely anyone does collections. That's the gap.
Invoice becomes overdue
→ Agent reads the full conversation history
→ Picks the right tone (polite → firm → final)
→ Drafts a personal follow-up — you approve
→ Customer replies (real email or simulation mode)
→ Agent parses: promised / disputed / question / ignored
→ Updates invoice status, tracks the promise
→ Schedules a re-check for the day after the promise
→ Still unpaid? Escalates the tone, drafts again
→ Stripe payment confirmed → done
That loop — read context, draft, approve, parse reply, escalate — is what makes this an agent, not a mail merge.
The hackathon asks for agents that earn, spend, and run real operations. Cashflow Agent does all three:
| How | |
|---|---|
| Earn | Collects overdue revenue — every paid invoice is money the agent recovered |
| Spend | Pays $0.25 per email send — recorded as a transaction, visible on the dashboard |
| Run | The collections loop runs autonomously — draft, approve, parse, track, escalate, repeat |
The dashboard shows Earned / Spent / Net in real time. The agent is self-funding.
| Layer | What it does |
|---|---|
| Hermes Agent Skill | 5 Python scripts: read thread, draft follow-up, parse reply, create payment link, schedule follow-up |
| Hermes CLI | Scripts call hermes chat -q for LLM reasoning (tone selection, reply parsing, customer simulation) |
| Hermes Cron | Scheduled re-checks fire automatically — "Check if invoice was paid by the promised date" |
| Next.js Dashboard | The user interface — cashflow board, agent actions, thread drawer, promise tracker |
The user never needs to know Hermes exists. They click "Draft" → the agent reads the full conversation, picks the right tone, writes the email. They just hit approve.
- Reply-aware follow-ups — the agent reads what was promised, what was said, and writes a follow-up that references it
- Reply parsing — paste a customer reply, the agent classifies it (promised / disputed / question / ignored) and updates the invoice
- Promise tracker — countdown to each promise date, urgency color coding (broken / critical / soon / ok)
- Tone escalation — polite → firm → final, based on days overdue and broken promises
- Batch collections — "Run Collections" drafts all overdue invoices at once
- Payment links — Stripe payment link per invoice, included in follow-ups
- Simulation mode — the LLM roleplays as the customer and auto-generates a realistic reply — the full loop plays out in 20 seconds, perfect for demos
- Collection metrics — collection rate, avg days to pay, promise-keep rate, escalation rate
- Agent economics — earned vs spent vs net, live on the dashboard
- Real email — Resend integration sends actual emails; inbound webhook auto-matches replies to invoices
- Google sign-in — one-click OAuth
- Demo mode —
?demo=1works without login, Stripe, or email keys
| Layer | Tool |
|---|---|
| Frontend | Next.js + Tailwind CSS |
| Database | SQLite via Prisma |
| Auth | NextAuth (Google + email) |
| Resend (outbound + inbound) | |
| Payments | Stripe Connect + Payment Links |
| Agent | Hermes Agent |
git clone https://github.com/getaskclaw/cashflow-agent.git
cd cashflow-agent
npm install
cp .env.example .env # add your keys (or just use demo mode)
npx prisma db push # create tables
npx tsx prisma/seed.ts # load demo data
npm run dev # → http://localhost:3000/dashboard?demo=1No keys? No problem. Demo mode (?demo=1) works without Stripe, email, or login.
Cashflow Agent 是帮你把钱收回来的 AI 助手。
点一个按钮，后台的 AI 代理读完你和客户的全部聊天记录，判断该用什么语气，写一封个性化的催款信，再自动安排下次跟进。你确认一下就行。
代理会赚钱（追回逾期款）、花钱（付邮件发送费）、自己运行（催收闭环全自动）。QuickBooks 只会群发提醒，Cashflow Agent 记得你们聊过什么。
Built with 🜂 by AskClaw