Written for the person who gets blamed if it goes wrong.
Eight plain statements. If one of them stops being true we change the page the same day.
Approval before action
Restart, deploy, spend, delete, rotate: each becomes a card in your channel that only a named owner can press. The agent waits. Read-only work, like diagnosing or reporting, needs no approval and is still logged.
A vault the model never reads
Secrets are stored by name. When a tool needs a token it fetches the value at the moment of use, server-side. The model sees "CLOUDFLARE_TOKEN", never the token. Nothing secret enters a prompt, a transcript or a memory.
One workspace per customer
Your agents run in their own isolated worker with their own limits and their own network rules. No customer shares a process with another. You can wipe yours without touching anyone else.
Your key, your cap
Inference runs on an Anthropic key you create in your own Anthropic console, with a spend limit you set. We never see your Anthropic bill and never carry it.
An audit you can export
Every job, every approval, every secret name used, with who pressed what and when. Download it any time as CSV. It is yours.
Hosted in the EU
Frankfurt, on servers we run. We are a data processor under UK and EU rules; a short DPA is available on request before you sign.
What we cannot see
Your secret values, your Anthropic bill, and anything in channels you have not invited an agent into. Collaborators only see what you allow.
What is not built yet
Slack and Teams arrive after Discord. The web pane for approvals arrives in October; until then approvals live in your channel. We say so on every page.
We set up your first agent with you.
One call, an afternoon of setup, thirty days with us in your channel. A human replies the same working day.
First month, we set it up with you. Not useful by day 30, full refund. Cancel any time.
Prefer email? hello@kamobo.co.uk