Connected inboxes
Plug your wholesale and support mailboxes into the workspace. Mail lands next to the customer record, the assistant drafts replies you approve, and nothing lives in someone's personal inbox.
AI for roasters
AI built into the roastery, not bolted on.
Roastnet ships an agent platform underneath every workflow. It plans roasts, forecasts demand, drafts invoices, replies to wholesale inquiries, and turns cupping notes into structured data. Opt-in by default.
AI roast planner
A roast schedule that drafts itself.
The planner reads open orders, subscription demand, packaged stock on hand, and historical roast patterns. It drafts a schedule for the day or the week and hands it to the operator to approve. Confidence threshold and monthly budget are configurable. The AI never roasts behind your back. It proposes; you approve.
- Inputs: open orders, standing wholesale deliveries, subscription cycles, safety stock.
- Outputs: an approvable plan, not an executed plan.
- Configurable confidence threshold to suppress low-quality suggestions.
- Monthly budget cap so the spend is bounded by you, not the model.
Demand aggregation, sharp enough to roast against.
Behind the planner is a demand-aggregation service that turns the messy real world into a single signal. It rolls up orders, subscriptions, contract obligations, and pop-up bookings. The AI never sees raw rows; it sees an aggregated view it can reason about quickly and cheaply.
Agent platform
Every extension exposes a toolkit. The AI can act, not just answer.
Roastnet runs on the agent platform: a router that picks the right tools, an orchestrator that decomposes complex requests, and a master-slave worker pattern that keeps cost down. Every extension ships a focused toolkit, so the AI is grounded in real, permission-aware operations.
- Router selects 2–4 focused toolkits per request, not a 70-tool buffet.
- Orchestrator decomposes complex prompts and spawns slave workers.
- Per-user permission filtering at spawn time, job time, and execution time.
- Every action logs to AI usage with model, tokens, and cost in cents.
Beyond chat
Agents that run the busywork, not just answer questions.
The assistant is one surface. Underneath it, the same agent platform watches your inboxes, reacts to events, works the phones, and coordinates teams of workers, all permission-aware and credit-metered.
Listeners
Event automations without code: when an order lands, a lease is signed, or stock dips, a listener fires and runs the follow-up: notify, create, escalate.
Voice agents & dialer
Outbound calls with a governed AI caller for order confirmations and wholesale follow-ups, with transcripts and call logs on the record.
Agent teams
For bigger jobs, a master agent decomposes the work and runs focused workers in parallel, then reports back with one result you can act on.
Real tasks the AI can take off your plate.
🗓️
Plan the roast day
Drafts tomorrow's schedule from real demand. You approve or edit.
📈
Forecast the next month
Reads order history and subscription growth to project bag-level demand.
✉️
Draft wholesale replies
Auto-drafts an answer to a café's order question. Grounded in their order history.
🧾
Invoice and chase
Writes the invoice, sends it, and chases overdue payments on the schedule you set.
🫘
Turn cupping notes into data
Structures free-form tasting notes into scores, descriptors, and faults.
📦
Flag stock risk
Watches the schedule against on-hand green and roasted stock. Raises alerts before you run out.
Transparent AI billing
Prepaid credits. A visible meter. No surprise bills.
AI usage runs on prepaid credits with a live usage meter, so you always know what an agent action costs before and after it runs. Every AI feature is off until you switch it on, and the usage log ships on every plan.
Opt-in by default. Visible always.
Every AI feature is off until a roastery turns it on. The credit usage log is on by default and never hidden behind a tier. The agent never sends roastery data to a third party other than the model you have chosen. There are no surprise calls.
On the roadmap
Crack detection. Vision-grade cupping. More on the way.
We are training a small model on Roastnet curves to flag crack events more reliably than the current heuristics. Vision-grade screen reads for chemistry sheets and producer cards are next. The agent platform is the substrate. New features plug into it without rewiring anything underneath.
Try the AI on a real roast day.
Start a free trial, connect your green book, and let the planner draft tomorrow's schedule. Approve it, edit it, or ignore it. You stay in charge.