mentat

native buzz support

one person knows why the retry logic backs off that way

Nobody wrote it down, and nobody is going to. Mentat joins your Buzz workspace and answers when you mention it. When the answer is in none of your sources, it interviews the person who holds it.

create a workspace →

bring one channel and a question you already know the answer to

what mentat is

A knowledge sharing layer for engineering teams. It answers questions against your code, docs, PRs and issues. Where nobody ever wrote the reasoning down, it runs a short interview with whoever holds it. Facts are easy to write down. Reasoning never is, and reconstructing it is the part Mentat does.

what runs on buzz

Four capabilities over one relay connection and one Mentat identity. Mentat is a member of your workspace, not a separate app your team has to switch to.

answers
Mention Mentat in a channel it watches and it replies in the thread, with the sources it drew on listed underneath. Anything that doesn't mention it starts no answer, because the trigger filter runs before the model call rather than after.
history
Channel history becomes part of what Mentat can answer from, next to your repos, docs and tickets, and you choose which channels it reads.
interviews
Mentat's interviews run over encrypted Buzz DMs (NIP-17), on the same interview flow as the web app, Slack and Teams. An admin links a member's Buzz identity, and that link only counts as verified once the person answers from the key they claimed.
git
Patch, PR and issue threads on Buzz (NIP-34) become a source like any other, for the repos you list and no others. Attribution comes from the key that signed the event, so a committer line someone typed by hand can't claim work it didn't do. Threads are correlated against the areas of your system they touch.

when it doesn't know

When retrieval turns up nothing to ground an answer in, Mentat posts one sentence and stops rather than synthesizing from thin context.

An answer in a channel is read by people who never asked the question.

"I don't have enough captured knowledge to answer that confidently yet."

the exact string it posts, not a paraphrase of one

An answer built from your sources arrives with them listed, so you can go read them and disagree.

getting connected

An admin connects Mentat from the workspace settings. The parts worth knowing before you do.

relay
You supply the URL Mentat connects to.
channels
You supply the channel IDs it watches.
keypair
Mentat generates its own and hands you the npub. You never give it a private key.
identity
Stays the same across reconnects, restarts, and a disconnect. Mentat never rotates it.
profile
A kind:0 profile goes out on every connect, so the mention picker shows Mentat rather than a raw npub.
scope
Watched channels and watched repos are both lists you set. Nothing outside them is ingested.

Mentat runs multi-tenant by default. It also runs single-tenant inside your own AWS account, where your data and your knowledge graph stay in your cloud and model inference still calls out to the provider. The security whitepaper lists every egress destination.

try it against something you already know

Connect one channel and ask it a question you can grade yourself. That is a harder test than anything a slide could show you. If the answer comes back thin, the sources it used are listed underneath it, so you can see exactly what it had to work with.