We run hotato inside your VPC.
The same MIT build you can download, deployed into your cloud account and operated by the people who maintain it. The audio and the evidence stay in your boundary; the upgrades and the pager are ours.
One address, answered by a maintainer: start the scoping email →
Nothing has to leave your network for this to work.
A hosted QA platform needs your call recordings in its account. Hotato does not, because there is no hosted backend to send them to.
The workspace binds 127.0.0.1 and publishes one port. The registry, the evidence store, the bearer token and the audit log all live in one volume you own. Everything the scorer needs is inside the compose network, so the deployment can sit in a subnet with no route out and still do its job.
That posture is the product of the open build, so you do not have to take our word for how the managed deployment behaves. ./deploy/verify-zero-egress.sh re-runs the image against its assertions, and you can run it yourself against the deployment we hand you.
The posture, command by command: security → · self-host →
- Cloud account
- Yours
- Audio at rest
- Your storage
- Evidence store
- Your volume
- Workspace binding
- 127.0.0.1:8321
- Codebase
- MIT, unmodified
- Operated by
- Attention Labs
What the agreement covers.
Operation, not capability. Every check that decides a verdict is already in the build you can read; what you are buying is somebody accountable for running it.
- Deployment into your VPC.
We stand up the compose deployment in your cloud account, sized to your call volume, and wire it to the storage and network you nominate. You keep the account and the credentials throughout.
- Capture wired to your stack.
We connect the adapter for the stack you record on and confirm it hands back one party per channel, which is what makes an overlap attributable. Vapi, Twilio, Retell, LiveKit and Pipecat are dual-channel today; Bland, ElevenLabs, Synthflow, Millis and Cartesia are mono, and we will tell you what that costs you in attribution before you sign anything.
- An availability commitment.
Written into your agreement and agreed per deployment. The figure follows from the region, the redundancy and the recording volume you run, so we set it against your deployment rather than printing one here.
- A named engineer.
One maintainer assigned to your deployment and reachable directly. They read the failing calls with you, and the fix lands in the public repository where you can read it.
- Upgrades and evidence continuity.
We take the version bumps, and the
/datavolume carries across them: a rebuild upgrades the image and leaves the registry and evidence untouched. The store is content-addressed with sha256, so a restored artifact is the same bytes that produced the verdict. - Governance you can hand to a reviewer.
We operate the controls the build already ships: bearer-token authentication on every workspace request, an audit log of those requests in your volume, and Ed25519 signing of evidence through the
[sign]extra, which verifies from the public key alone. We configure them, keep them on, and produce the records when your compliance review asks. - The CI gate, running against your suite.
We set up the composite GitHub Action or the pytest plugin against your pinned calls, so a regression fails the build rather than reaching a customer. The step exit is hotato's exit code: 0 pass, 1 fail, 2 refuse.
What is not in this listAnything that would take a capability out of the free build. The open core stays MIT and complete, so a self-hosted team measures exactly what a customer of ours measures. The free tier, in full →
Send one email and we will scope it.
There is no form and no qualification funnel. A maintainer reads the address below and answers it.
Tell us what you record
The stack, the monthly call volume, the cloud account and region, and the availability you need. Six lines is enough to price it.
Bring one bad call
Run hotato autopsy on a call that went wrong before you talk to us. The scoped conversation is far shorter when we are both reading the same scorecard.
Email [email protected]
The subject and the six questions are prefilled. Replace the blanks and send it.
Open the scoping email →The same address takes security reports, per the security page and the package's own SECURITY.md.
What a procurement review usually asks.
Answered against what the build does, so your reviewer can check each one in the repository.
Where does our audio live?
In your storage, in your cloud account. Scoring runs on the machine that invokes it and reads local files. The one off-box audio path is a hosted diarizer behind an explicit --egress-opt-in flag, and the default diarizer is local, so it is a choice you make rather than one you inherit.
Does a model see our calls?
Only if you turn one on. The deterministic lane decides the verdict without calling a model at all. The judge is a separate advisory lane, its verdicts are quarantined from the deterministic count, and in the self-hosted stack the model runs on Ollama inside the compose network with no published port.
What is the availability commitment?
Agreed per deployment and written into your agreement. It depends on the region, the redundancy and the recording volume you run, so a number printed on a marketing page would be the wrong number for most readers.
What happens if we stop paying?
You keep running it. The deployment is the MIT build in your own account, so the end of an agreement returns you to self-hosting rather than switching anything off. The /data volume with the registry, evidence store, token and audit log is already yours.
Can we read the code that decides a verdict?
All of it. One MIT repository holds every check, and the fixes your named engineer writes land there in public. github.com/attenlabs/hotato →
Try it first. Then tell us what you want operated.
The build we would deploy for you is the one you can install right now.
$ pip install hotato
Scope a deployment, or read what the free build covers. [email protected] → · Pricing →