Retrieval is the easy half
Which is why most assistants stop at answering. "Here are your three open tasks about the redesign" is safe. "Done, I closed it" is not — unless closing it can be taken back.
So the design question is not how many actions the chat can perform. It is what happens in the seconds after it performs one you did not want.
The four things it can change
MetaChat's mutation set is deliberately small:
That is the whole list. It is not short because more was hard to build; it is short because every additional verb is another way to quietly damage a record you rely on. Anything outside the set gets answered rather than executed.
Undo is the actual feature
Sixty seconds is a deliberate compromise. Long enough to read what happened and react. Short enough that the button is not still sitting there tomorrow next to an action you long since accepted, inviting a confusing reversal.
The important property is that undo is scoped to the action, not to the conversation. Reversing one thing does not unwind everything else that happened in the same session.
Why it will not run on a local model
This is the limitation most likely to matter to you if you came for the privacy story, so here it is plainly.
A local model asked to close a task will sometimes emit something that looks like a tool call and is subtly wrong — the wrong task, malformed arguments, a call when it should have answered in prose. In a chat that only retrieves, that is a bad answer. In a chat that mutates, it is a wrong write.
So the feature declines rather than degrades. Everything else in MetaWhisp — dictation, transcription, extraction — still runs on device or on a local model. This one part does not, and the app says so rather than shipping something unreliable and calling it support.
There is a difference in reliability even above that line. With Pro the model returns a structured list of tool calls natively. Without it, the tool call is parsed out of the model's text output, which works but is inherently more fragile. Both go through the same confirmation and undo path.
What it knows
The chat reads from whatever the second brain has collected and you have enabled: extracted memories, tasks, transcripts of your dictation and calls, and screen context if you switched that on.
That conditional matters. Memory collection is off by default. Screen reading is off by default. If you have not turned them on, the chat has correspondingly less to work with — which is the right trade, but worth knowing before you conclude it has a poor memory.
Where this sits next to a general assistant
| General AI chat | MetaChat | |
|---|---|---|
| Knows your week | Only what you paste in | What you dictated and discussed |
| Can change your records | No | Four scoped operations |
| Undo | Not applicable | 60 seconds, per action |
| Runs on a local model | Sometimes | No — needs reliable tool calling |
| Best at | Open-ended thinking, writing, code | Acting on your own record |
These are not competitors. A general assistant is better at almost everything except the one thing it structurally cannot do: know what you agreed to on Tuesday and close the task that came out of it.
