Closed beta · free of charge

Product support that never makes things up.

The Support Assistant is an embeddable help widget that answers customers from knowledge you have approved — and nothing else. When it does not know, it says so, and hands you the question to answer once, publish, and keep.

No credit card, no tracking. This page loads no third-party scripts, fonts, or cookies — neither does the widget on your site.

How it works

Teach it once. Publish deliberately. Stop repeating yourself.

Three steps between "no support" and a widget you can actually trust in front of customers.

Embed the widget

A launcher you host opens a sandboxed, cross-origin iframe — lazily, only when a visitor asks for help. None of our JavaScript ever runs in your page.

your page → your launcher → sandboxed iframe

Teach it what is true

Use the owner console to write answers directly, or let scans of your Git repository and Notion pages suggest the questions worth answering first.

evidence → suggested questions → your drafts

Publish deliberately

Nothing becomes a customer-facing answer until you explicitly publish it. Questions the assistant cannot answer queue up for you to teach — once.

draft → explicit publish → live answer

What you get

A support stack, not just a chat bubble.

Answers

Citations, checked server-side

Every answer cites the published knowledge it came from, and the citations are verified before they reach the visitor.

Honesty

"I don't know yet" built in

Instead of improvising, the assistant records a knowledge gap. Your unanswered questions become a prioritized teaching queue.

Console

Private preview chat

Test your published knowledge as a public visitor or a signed-in user before customers ever see an answer.

Sources

Evidence import, never auto-answers

Git and Notion scans surface what your product can do — as suggestions for you to answer, never as text published on their own.

Offers

Deterministic help offers

Proactive tips are matched by explicit rules you configure — no model deciding when to interrupt your users.

Scale

One deployment, many products

Each product gets its own knowledge base, allowed origins, audiences, and answer queue under a single owner console.

Trust model

The model may phrase approved material. It never decides what is true.

For the developer who has to sign off: this is the entire path from your product knowledge to a customer answer.

your evidence   git / notion scan ──▶ suggested question ──▶ open gap ──▶ owner draft ──▶ explicit publish

customer question ──▶ published knowledge retrieval ──▶ constrained answer + checked citations
                  └──▶ insufficient evidence ──▶ "not known" + open gap for the owner
  • Sandboxed by construction. The widget is a cross-origin iframe with a strict Content Security Policy, embeddable only by origins you allow.
  • No third-party script in your page. The launcher is yours; the iframe loads lazily. Your page never executes our code.
  • Signed-in context is signed. Product state arrives as short-lived Ed25519 tokens, consumed once — the widget never sees your users' credentials or data.
  • Audience boundaries are server-side. Public visitors and signed-in users draw on separately scoped knowledge, enforced at retrieval.
  • Metadata-only accounting. Model usage is tracked without storing prompts; transcripts expire on a retention schedule you configure.
  • Abuse controls by default. Rate limits, session caps, and message limits are on out of the box, with pseudonymous client accounting.

FAQ

Fair questions.

What does it cost?

Nothing. The Support Assistant is in closed beta and free of charge. If that changes, beta participants hear about it first, with plenty of notice.

Can it answer from our docs automatically?

Deliberately not. Imported Git and Notion material is evidence for suggesting questions — a person approves and publishes every answer. That is the feature, not a limitation.

What happens when it does not know?

It says "I don't know that yet", records the question as a knowledge gap, and shows it in your answer queue. You answer once — directly or with a guided teaching conversation — then publish.

What powers the answers?

A small, fast language model phrases answers from your published knowledge, with retrieval over PostgreSQL and pgvector. The model cannot publish, browse, or invent product facts.

Who is already using it?

MemoryTether, a portable-memory service, is the first project configured for The Support Assistant. Its support knowledge is being prepared before the widget is integrated live.

What do you store about my customers?

Chat transcripts with a configurable retention window, pseudonymous rate-limit counters, and metadata-only model accounting. No tracking scripts, no cookies on this site, no third-party requests from the widget.

Closed beta

Ask to join — it is free.

Tell us where you would put a support assistant, and we will set you up with your own knowledge base and widget, free of charge during the beta.

We use your email only to reply about the beta. Requests land in a queue reviewed by a person.