Typeless, translated through again: no hard proof on the model, but I dug up the search pipeline

AIYA
, , ,
1

This time I went through the local Typeless on my machine like a little detective project. The goal was straightforward: what model does it actually use? Why does its “press once to listen to you, then reply directly” interaction feel so smooth—and why, specifically, can’t you do multi-turn follow-up questions?

First, the conclusion version:

  • I didn’t find a single hardcoded model name in the local client. I didn’t see any straightforward strings like gpt-4o, claude, or gemini.
  • It’s not a local model. The core requests go through https://api.typeless.com, and the main AI entry point is /ai/voice_flow.
  • Its “search” capability is real, and I captured a record in local history containing web_metadata, where a grounding redirect link to vertexaisearch.cloud.google.com appeared directly.
  • But Typeless’s official public pages have also said it uses third-party LLM providers, such as OpenAI. So it’s more like “the server switches capabilities by scenario,” rather than the client being tied to one model from start to finish.

The evidence I actually dug up is pretty interesting:

  • Install location: C:\\Users\\1\\AppData\\Local\\Programs\\Typeless\\Typeless.exe
  • User data: C:\\Users\\1\\AppData\\Roaming\\Typeless.exe
  • Hotkey setup is exactly what I usually use: RightAlt to record, RightAlt+Space for hands-free mode
  • The local database has only one core business table: history
  • I didn’t see multi-turn conversation tables like conversation, thread, or message
  • In voice_command records it mainly stores user_prompt, refined_text, web_metadata, external_action, and delivery

This also conveniently explains the two things I both love and hate about it:

  1. The newly popped-up chat window basically doesn’t have the previous context.
    The reason is simple: it’s more like a “one-off voice command,” not a persistent chat thread.

  2. After it replies, you can’t naturally continue with follow-up questions based on the previous one.
    Because structurally, locally, it isn’t designed around multi-turn conversations—more like “you say one thing, I give one result, and this transaction ends.”

There’s also a very key piece of hard evidence:

When I asked a real-time question like “Thor Accelerator was maliciously attacked—what was the attacker’s goal? Search it,” the locally stored mode_meta.ai_result.web_metadata included:

  • grounding_chunks
  • grounding_supports
  • vertexaisearch.cloud.google.com/grounding-api-redirect/...

And in the sources you can see sites like sina.com.cn, cnyes.com, and youtube.com.

This shows that at least for some “real-time info” questions, it really does go through a grounding workflow with search sources, rather than just making things up on the spot with a serious face.

So the most reliable takeaway this time is:

Typeless is more like a “voice-enabled global AI command panel.”
What it’s good at: fast, smooth, callable from anywhere, and combining the current UI context to give a direct result.
What it’s not good at: keeping long context, and doing repeated multi-turn back-and-forth around the same answer.

In other words, it’s not a “voice ChatGPT chatroom,” but a “voice-driven global AI quick-action layer.”
That’s exactly why it feels good to use—and also why it makes people want to complain, “Why can’t I keep chatting?”

If you like this kind of interaction, I also quickly looked at similar products:

  • Wispr Flow: the closest to Typeless; more “global hotkey + input into any app”
  • Superwhisper: great for “select text and then rewrite/ask something in one sentence”
  • Aqua Voice: heavier on context awareness
  • Braina: more like a classic Windows voice assistant plus an AI writing assistant

One-sentence wrap-up:

Typeless isn’t incapable of chatting—it simply wasn’t designed as a “chat tool” in the first place; it’s more like an on-call voice AI helper that gets to work after you speak and then wraps up.

If I keep digging deeper, the two things I most want to focus on are:

  • Whether the debug fields returned by the server can be further decoded into more specific provider clues
  • Which questions trigger search and which ones go through plain generation only
1 Like