Answering through Claude

Whispr can route detected questions to a Claude session you are already in, alongside the provider engines rather than instead of them. Claude answers with your skills, your files and your other MCP servers behind it — which is the whole reason to want it: a provider engine sees the question and your prep, and a Claude session on your machine sees everything else you have given it.

It is optional. Nothing here is needed for Whispr to work.

Two different things both called Claude

This is the confusion worth clearing up first, because the app shows both.

They are independent. You can run either, both, or neither, and switching one off does not affect the other.

Setting it up

Everything is in Settings → Claude as the brain. You do not need a terminal, and there is no whispr command to find — the buttons ask the daemon to do it.

1. Tick the box to route questions to an attached Claude session. If no brain engine exists in your config yet, ticking it adds one.
2. Give Claude the tools registers this daemon as an MCP server, at user scope so it works in every project rather than only the folder you happen to be in.
3. Install the skill copies the answering loop into ~/.claude/skills. Without it Claude can see the tools and does not know what to do with them — which presents as a session that connects and never answers.
4. Restart Claude, then tell it to start answering Whispr questions.

Step 4 is the one people miss. MCP servers are read when a session starts, so a session that was already open when you pressed the buttons will not see Whispr at all.

The panel then says ✓ a Claude session is attached, with a running count of answered, declined and waiting.

If the button cannot do it

The first button shells out to the claude CLI. If you use Claude Desktop, or the CLI is not on your PATH, it says so and nothing is broken — the exact command is printed below the buttons with a Copy button, and running it yourself does the same job. The command carries this daemon's real port, so copy it rather than typing it from memory.

What Claude can reach

Once attached, a Claude session can use these tools. They are the same ones the skill teaches it to drive:

• the answering loop — claim a detected question, stream an answer back into the card, settle it
• search_knowledge — your prep material, scoped to the meeting you have selected
• get_meeting_brief — what this meeting is, what kind it is, and the generated reviews that are not otherwise retrievable
• add_prep_material — put a document into prep from the Claude side
• the history and practice tools, for reviewing a meeting afterwards

That last point is worth knowing before a call rather than after: you can prepare through Claude — ask it to research the company and file the notes — and the same material grounds the live answers, because it went into the same place the Prep screen writes to.

What it costs, and what it is slow at

The brain runs on a Claude session you are already paying for, so it adds nothing to your provider bill. What it costs is time: it gets 120 seconds per answer against 60 for the others, because a session reads your prep and thinks before the first token. It is the slowest engine you have, deliberately, and it answers beside the fast ones rather than holding them up.

It will never be your classifier. The brain answers questions; deciding what was a question stays on a small fast model, and both surfaces filter the brain out of that role explicitly.

When it does not answer

Nothing attached. The panel says so, and questions sent to the brain will time out unanswered while every other engine keeps working. Restart Claude and tell it to start answering.

Whispr has to be running. The MCP endpoint is served by Whispr's own daemon, so if Whispr is not running, claude mcp add still succeeds and the server simply never connects — which looks like Claude ignoring you.

What counts as "running" differs by platform, and only for audio. On macOS the daemon runs headless: whispr start with no window serves the MCP endpoint, captures the call and answers, so Claude can be attached to a machine with nothing on screen. On Windows the audio is captured inside the app window, so a headless daemon serves everything here — the tools, your prep, the history — and hears nothing. Either way, the window being open is always enough.

"No prep material covers this" is Claude declining on purpose, not a failure. It is counted separately from an answer, and it means what it says: add the material, or let another engine take that one.