One answer each, then a way to reach a human.
The eight things people ask most, kept short. If yours is not here, the form at the bottom reaches someone who can read your run output.
Install and first run
Install, then point the engine at a directory. The first run analyses and reports; nothing is changed on disk until you ask it to apply fixes.
Licence keys and activation
You generate the key yourself in the account portal, and it is returned to you there rather than emailed. Activating issues a key for your current plan and revokes the previous one in the same step. The engine checks it offline against a public key compiled into the binary.
Go to the portal →Billing, seats and plan changes
Pricing is per seat and the count is the quantity on the subscription, changeable in the portal at any time. Invited team members inherit the plan without a subscription of their own. Plan changes and invoices go through the billing portal from your account page.
See the plans →CI and coding assistants
CI is a workflow step, and the engine registers with your coding agent in one command. Both are in the free tier.
Privacy and what leaves the machine
Detection needs no account and no API key, and no source leaves the network. A paid plan does mean an account, for licensing rather than analysis. If local rewording is enabled, the only thing sent to the model is one already-generated English sentence, and the model runs on your own hardware.
How we measure →Reporting a false positive
Open an issue on the engine repository with the finding, the file it fired on, and why it is wrong. Adjudication happens in public, and detectors that lose an adjudication are quarantined with a recall test rather than quietly deleted. Published precision is per detector for exactly this reason.
Open an issue →Contacting a human
Use the form below. Version, platform and the run output get you a real answer on the first reply instead of the third.
Go to the form →Setup, and the five things that go wrong.
Every failure in this feature is quiet by design: missing weights, a disabled setting, or an unreachable server all fall back to the original sentence. One command says which it is.
On Linux with an NVIDIA card the prebuilt server is CPU-only. A GPU build has to be compiled locally, and the guide in the engine repo does it without root.
It reports what is missing and the single next command to run. Nothing else in this feature announces its own failure, so this is the first thing to try and usually the last.
Returns verified prose for a finding.
Reports setup read-only, and hands back a CLI command rather than downloading gigabytes through a tool call.
I enabled it, but the wording never changes.
Run truemend model-status. Usually the server is not running or the weights are absent. It is also expected most of the time: about one sentence in five is reworded, and the rest are returned unchanged. An unchanged sentence is the canonical sentence, so nothing is lost.
Is my code sent anywhere?
No. The model runs locally, and it receives one already-generated English sentence. Never source, never findings, never counts. An address that is not local triggers an explicit warning, because that would mean text leaving the host.
Why is it slow?
About 1.64 seconds a sentence on CPU against 0.26 on GPU. On Linux with an NVIDIA card the prebuilt server is CPU-only and a GPU build has to be compiled locally, so latency is the only symptom you will see.
The download failed a checksum.
The file is quarantined rather than installed, which is intended behaviour and not a bug. Re-run the download. If it keeps happening, the mirror is serving something that does not match the pinned hash, and that is worth reporting.
Can I use a bigger model?
One model ships and the command line accepts only that one. Others exist in the registry for benchmarking, not for use.
Send the run output. It saves a round trip.
Version, platform and the output of the run that went wrong answer most questions before anyone has to ask. Paste it verbatim, redacted if you need to.
A suspected false positive is better as an issue on the engine repo, where the adjudication happens in public.
Open an issue →