Billing desk
Answers invoice questions with the account's history, drafts the reply, waits for approval.
Billing deskrun #4,218 · $0.004
09:41:02Billing desk woke on message received in #billing
09:41:02Read 1 message, 1 attachment (invoice-2041.pdf)
09:41:03Looked up account · found 2 prior invoices
09:41:04Drafted reply · permission send without approval is off
09:41:04Waiting for approval · cost so far $0.004
Hi Daniel — invoice #2041 covers the March site visits. Your last two invoices (#1987, #2010) are paid in full, so nothing is outstanding beyond this one. Happy to split it across two months if that helps.
Approve & sendEditDiscard
Sent · $0.004
The designer
Configured, not hardcoded
Agents are global objects you assign to any number of channels. Edit one and it changes in every lane it works. Each has a visual identity, a purpose, skill files, permissions, a budget and a model.
- Drag onto a channel the agent bar on each channel is where assignment happens.
- Per-agent credentials your key or Sodaroad-managed, chosen per agent.
- Budgets monthly limit, per-task ceiling, and pause or notify at the limit.
The seven permissions
- ✓Read messages See what arrives in the channel, and nothing outside it.
- ✓Draft replies Write the reply and leave it on the row with Approve and Edit.
- ✓Send without approval Off by default. Turn it on for a lane you trust it with.
- ✓Archive and label File what should be filed, and say so on the row.
- ✓Access attachments Open the invoice, the photos, the PDF.
- ✓Call external tools Look things up in the systems you connect.
- ✓Contact new recipients Off means the agent can only reply to who wrote in.
Approval is on the row
There is no queue to check.
Agent output renders inline on the message row in three states: reply drafted, needs approval with Approve and Edit right there, and archived by agent. A filter shows everything that came from an agent.
3
states on the row
1
enforcement point
7
permissions
$
budgets in dollars, not tokens
Use cases
What people put agents on.
Triage
Reads a busy channel, archives what is noise, flags what needs a human today.
Scheduler
Turns 'can we find a time' into a proposed slot, with your calendar as a tool.
Vendor replies
Auto-replies on a lane you trust it with, under a recipient allowlist and a rate cap.
Newsletter reader
Summarises the week's issues into one message in the channel.
Support intake
Acknowledges, asks the two missing questions, files the rest for a person.
FAQ
About agents.
Which models can an agent use?
Any. Bring your own API key (OpenRouter first, Anthropic and OpenAI direct) and the agent runs on your bill, or use Sodaroad-managed models at cost on a paid plan. The choice is per agent.
How does an agent know when to act?
An agent has a free-text purpose plus uploaded skill files. Triggers live in the skill file, bound to Sodaroad's event schema: message received, thread updated, mention, reply awaited, schedule elapsed, and more.
What stops an agent from being tricked by a malicious email?
Permissions intersect with the channel's scope at one enforcement point; 'contact new recipients' off is a hard recipient allowlist; rate caps and a loop breaker sit on every send; every action is audited; and each channel has a kill switch.
What does an agent cost?
What the model charged us, no markup. Each agent has a monthly budget and a per-task ceiling in dollars, and the live cost shows in the activity tray while a run is happening.
Put an agent on it.
Chat and mail are free. Agents start at $10 a month, billed at cost.