For engineering leaders
Your team could build this.
Here's what it takes.
The AI is the easy part. Running it safely for a whole company is the work: sandboxes, approvals that actually hold, credentials, retries, logs and cost. We built that part, so your engineers don't have to.
- 0.0sStarted by schedule · fresh machine
- 1.8shostfully.get_reservations · 12 rows
- 4.1ssifely.create_code · Maple St
- 6.3sDraft ready · send held for approval
- 2m 10sApproved by Danny · sent in a separate step
Inputs, outputs, errors and timing for every call. Readable in the app and through the API.
What it solves
Nine problems you'd hit in the first month.
Each one is solved in the platform today. Ask us to show any of them live.
AI code runs next to your data
Every run gets its own fresh machine, and it's destroyed when the run ends.
Nothing carries over from one run to the next.
“Ask before sending” is a checkbox the model can skip
The run that writes the draft never holds the keys to send. A person's yes starts a separate step that sends.
Approval is where the send rights live, not a prompt instruction.
Jobs fail halfway, or run twice
Work sits in a durable queue with leases, retries and duplicate checks. Each Dozal remembers what it already handled.
A restart picks up the work; it never repeats a send.
Keys end up everywhere
Credentials are encrypted per company, each run gets a pass scoped to that run, and secrets are masked in every log.
Website logins happen in a live browser window a person controls.
Nobody can see what it did
Every tool call is recorded with its input, output, error and timing, plus a plain-words reason when a run fails.
Your own agents can read the same records to debug.
The bill is a surprise
Every run's cost lands in an append-only ledger, by Dozal and by run. Runs stop at cost and step ceilings.
A Dozal that keeps failing pauses itself and says why.
One client's data leaks into another's
Every query is scoped to the company. Ask for another company's record and the answer is “not found”.
Agencies run one walled-off space per client.
Prompts change and nobody knows who changed them
Every change to a Dozal's instructions is versioned with who made it, including which API key, and can be reverted.
A Dozal can propose its own edits; a person approves them.
You're locked into one model
Models are swapped behind the platform, with a fallback provider. Nothing on your side changes.
Your Dozals never depend on one AI company.
How a run works
Five steps, every time.
The same path for every Dozal, whether it answers a guest or writes a report.
Something starts it
A schedule, a Slack tag, a new email or booking, an API call, or another Dozal finishing.
A fresh machine
It gets only the access this job needs, through a pass scoped to this run.
It does the work
Reads your apps, drafts the result. Anything risky is held: sends, money, deleting, door codes.
A person decides
Send it, change it, or don't, from Slack, the phone or your own tools through the API.
A separate step sends
Only after the yes. Then the machine is gone, and the record and cost stay.
Your stack
Your agents drive it. Nobody leaves their tools.
Everything the app does is an API call with a published spec. Use the command line, or connect Claude, Claude Code or Cursor in a minute.
- Keys at Viewer, Member, Admin or Owner level
- Keys expire, can be revoked, and can't create keys or touch billing
- Test runs before anything goes live
- Calls to your apps listed by key name
$ npx dozaldevs login $ dozaldevs list-dozals Door Codes · Owner Reports · Review Watch $ dozaldevs get-task-events 48213 5 steps · 2m 10s · approved by Danny · $0.04 # or plug it into Claude Code $ claude mcp add --transport http \ dozaldevs https://app.dozaldevs.com/mcp
Straight answers
What we don't do yet.
You'd find these in week one anyway. Better to hear them from us.
- SOC 2 report
- Not yet. We send a written summary of what's in place.
- Single sign-on (SSO, SAML, SCIM)
- On the roadmap.
- Spending limits your team sets
- On the roadmap. Today: per-run ceilings, a prepaid balance with a low-balance warning, and pause.
- One exportable audit log
- On the roadmap. Today the trail is edit history, approvals, run records and key calls, all readable through the API.
- Self-hosting
- Not offered. Cloud only, on US infrastructure.
Decide with real numbers, in 30 days.
Run one real job with one team. At day 30 your engineers look at the run records, approval rate and cost per run, and make the call. Either side can stop there.
- Your own agents drive it from day one
- Nothing reaches a customer without a person's yes
- Everything a Dozal is and did stays readable through the API