The assistant does not answer

Start with what the screen said

The assistant reports its failures in words, under the face, and each message points at a different repair. Read it before changing anything.

• "You've used this month's assistant allowance." — the monthly allowance is spent. Everything else on the screen carries on working.
• "The assistant isn't included on your plan." — this plan has no assistant allowance at all.
• "This account is suspended." — nothing will be spent until the account is back in good standing.
• "Microphone unavailable — press T to type instead." — the microphone could not be opened. The assistant itself is fine.
• "I couldn't hear that — try again?" — audio arrived but nothing could be transcribed from it.
• "Something went wrong reaching my brain." — the turn started and failed part way through.

If there is no message at all — the face does not wake, nothing is transcribed, nothing is spoken — the screen is probably not connected. Check the last-seen time on the Screens page first.

The allowance

Every assistant turn is checked against the account's allowance before a single token is spent, so a spent allowance stops the assistant cleanly rather than half way through a sentence.

Open Plan & usage from the account menu. Assistant credits this month shows what has been used against what the plan allows, with a bar, and underneath it an estimate of how many more conversations that leaves. The allowance resets at the start of each calendar month.

1. Open Plan & usage

   It is in the menu under your own name, in the header.

2. Read the Assistant credits this month figure

   Used against allowed. If the two are level, that is your answer.

3. Decide between waiting and upgrading

   The allowance resets on the 1st. A larger plan raises it from the moment the change takes effect.

A web search costs about 3,000 credits, which is several ordinary conversations. An account that suddenly runs short after years of comfort has often just had searching switched on.

The microphone

The microphone belongs to the screen, not to Mirra, and it is the single most common cause of a mirror that appears deaf while being perfectly healthy.

The fastest test in the whole of this article: press T at the screen and type the question instead. If a typed question is answered and a spoken one is not, the fault is the microphone and nothing else — the allowance, the connection and the model are all fine.

1. Press the microphone button on the screen, or press M

   Both do the same thing. Pressing it again while it is listening stops it.

2. Watch for the permission prompt

   The first time a browser is asked for a microphone it asks the person in front of it. On a wall-mounted screen there is nobody there to say yes, so this has to be done once at setup.

3. If it was refused once, clear it

   A browser that has been told no does not ask again. Clear the microphone permission for the site in the browser's own settings and reload the screen.

4. Press T and type

   If typing works, stop here — the assistant is working and the microphone is not.

• "Microphone unavailable — press T to type instead." is what a refused or missing microphone produces.
• A screen with no audio output can hear you and cannot answer aloud. Run diagnostics on the screen's own page and it reports "no audio output" under What this screen can do.

The wake word

This is where expectations and behaviour part company most often. There is no wake word in a browser — on a television, a tablet, a mini PC — so saying the name at the screen will never do anything. Use the microphone button, or M.

The field is on the Assistant page, under Identity, labelled Wake word (Raspberry Pi). Its own hint says the same thing: it takes effect once the Pi agent is running, and browsers use the microphone button. The Pi agent in the repository is a skeleton — wake-word detection is listed there as a planned module rather than a shipped one — so the setting is stored and nothing is listening for it yet.

Wake word (Raspberry Pi)

The word the Pi agent is to listen for. Lower-case letters and spaces only, up to twenty characters — a value containing anything else is rejected outright when the field is saved, which is why a word with a hyphen or a number in it appears not to save.Default: mirror

Let the assistant search the web

Whether the assistant may look things up. It changes what the assistant knows, not whether it answers — but on an installation with no search provider configured, turning it on does nothing, and the page says so in amber above the switch.Default: Off

The connection to the assistant service

The assistant does not run on the screen. Speech travels from the screen to Mirra's assistant service over a connection the screen holds open, and the reply comes back the same way. If that connection is down, pressing the microphone looks like nothing happening: no transcript, no error, no face.

Screens reopen the connection themselves, waiting a little longer after each failure up to thirty seconds between attempts, so a screen recovers on its own shortly after the network does. What you cannot do is diagnose this from the screen — it shows no connection indicator. The last-seen time on the Screens page is where you find out.

• A screen that reads offline on the Screens page will not answer, and no assistant setting will change that.
• A screen that is online but silent is worth reloading once, from Screen commands on its own page.
• Home Assistant and WLED are reached from the screen across your own network, so the assistant can be perfectly well and still unable to turn a light on.

When replies are thin rather than absent

There is a failure that produces answers instead of silence, and it is the hardest one to spot. Mirra's assistant can run on Claude through Amazon Bedrock, on Claude through the Anthropic API, on Ollama on-premises, or — when none of those is available — on a small rule-based stand-in that handles simple requests and nothing else. The fallback is silent by design, so that a household keeps a working mirror rather than a dead one, and the only symptom is that the replies get noticeably thinner.

Whoever runs the installation can see this directly. Super admin, then the AI usage & cost tab, carries an Assistant status panel at the top that separates what the deployment asked for from what is actually answering right now. There is a Re-check button on it.

• Configured against Actually answering. When the two differ, the cloud provider refused and the panel shows the provider's own error, when it fell back, and when it will try the cloud again — it retries after ten minutes.
• "No AI provider is configured, so every reply is coming from the rule-based stand-in." — nothing has been set up to answer with.
• "The relay isn't answering." — nothing can talk to the assistant at all until that is fixed.
• A relay that is up but not reporting which provider it is using either predates the panel or has an INTERNAL_SECRET that does not match the portal's.

When the panel says the assistant is running on the configured provider, the plumbing is sound and any remaining complaint is about the prompt rather than the wiring. That is a different repair, and knowing which one you are in saves an hour.

Working through it end to end

1. Ask it something simple at the screen

   Read whatever appears under the face. Half the causes name themselves here.

2. Check the screen is online

   Screens, and the last-seen time. An offline screen explains everything below it.

3. Check the allowance

   Plan & usage, Assistant credits this month.

4. Press T and type the same question

   Answered by typing and not by speaking means the microphone, and only the microphone.

5. If you run this installation, open Super admin

   AI usage & cost, then the Assistant status panel at the top.

What it will not do, however healthy it is

• It cannot look anything up online unless web search has been switched on, and not at all on an installation with no search provider configured.
• While it is reading a web page it loses the ability to change anything in your home, on your screen or in your lists for the rest of that turn. It can answer, and that is all.
• Replies are deliberately short, because they are spoken aloud rather than read.
• It reads calendars, lists, your home and your screens; it does not create or edit calendar events.