Mira is 100% free software. Get it by direct download. The Microsoft Store version is available in Japan only.
One-click install · Auto-updates.
Trusted Microsoft distribution for effortless setup.
The Store version is available in Japan only. Everywhere else, please use the direct download below.
Version 2.3.1
Windows 10/11 (64-bit) · ~54MB · No .NET runtime required
Completely free · All features included
October 13, 2026
In October 2026 Microsoft moved its read-aloud service, so Edge TTS (the natural voice) stopped working and Mira fell back to the standard Windows voice. Mira now uses the new endpoint. (Edge TTS is used when Python's edge-tts is installed on your PC.)
Mira could pick up its own voice from the speakers, take it as you speaking, and keep interrupting and answering itself. The microphone is no longer sent while Mira is speaking or using a tool such as checking the screen (so you can't interrupt Mira by voice mid-sentence; saying the stop phrase still ends Live mode). Mira also used to wait until the whole reply had arrived before speaking, leaving several seconds of silence in which a small noise could cancel the reply. Replies now start playing as soon as they arrive.
If you started speaking right after Mira finished, the first part was not sent, which caused mishearing. This is now much less likely. Also, if you are still talking when a reply arrives, Mira now waits until you finish.
Live mode could take a nearby conversation as a request and look at your screen without being asked. Its instructions now tell the AI to respond only when you talk to it and to look at the screen only when asked. (As this relies on the AI's judgment, it may not be prevented completely. If you don't want your screen sent, turn off Screen Capture in the Live mode tab of Settings.) Screen Capture is now also off by default for new installs (existing settings are kept).
Short phrases could be misheard as another language, showing Spanish and other languages on screen or even getting a reply in them. In the Live mode tab of Settings, Language (listening & replies) lets you choose Japanese, English or auto-detect. The default is Match Windows language (auto-detect when Windows uses a language other than Japanese or English). With a language set, replies are in that language. With Japanese set, transcripts with no Japanese characters are shown as "…"; with English set, some transcripts that look Spanish or Portuguese are.
Surrounding sounds could be mistaken for "Mira live" and start Live mode by itself. Live mode now requires a higher-confidence detection than a regular wake word, so this happens less often.
Options that were on by default (Start with Windows, emotional expression, avatar, and Live mode PC control and screen capture) were not saved when turned off, and came back on at the next start. They are now saved correctly, and options you had turned off are switched back off when you start the updated version (unless you saved Settings after they came back on, so please check your settings once).
Google has replaced gemini-3.7-flash with Gemini 3.8 Flash, so selecting it already used 3.8 Flash. It has been removed from the list, and the setting of anyone who chose it now shows gemini-3.8-flash (the model actually used does not change).
The setup screen described the Gemini free tier as "up to 1,000 requests/day". Google does not publish a daily limit and the actual value varies by model and account, so this has been corrected. The screen now also mentions that on the free tier, what you send may be used by Google to improve its products.
October 10, 2026
OpenAI moves to the GPT-6 family (gpt-6.1-sol / gpt-6-luna), Claude to the Claude 5.5 family (Haiku 5.5 / Sonnet 5.5 / Opus 5.5), and Gemini Live to Gemini 3.8 Live. The default Claude model is now Haiku 5.5, at about one tenth of the price of Haiku 4.5.
The first time you start this version, any previous-generation model left in your settings is switched once to a successor at the same price or lower (for example gpt-5.6-terra → gpt-6.1-sol, Claude Haiku 4.5 → Haiku 5.5, Gemini 2.5 Pro → Gemini 3.8 Flash). v2.2.0 kept such settings unchanged; we changed this so you don't keep paying more or waiting longer on an older model. Names that always point to the newest model, such as gemini-flash-latest, and model names you typed yourself are left unchanged. You can still pick a different model in Settings afterwards.
The new Claude models can think before they answer, and when they did, screen recognition returned nothing. Mira now reads the answer correctly. If the screen can't be read, Mira tells you that the analysis failed instead of guessing.
Changing your API key or model in Settings did not reach screen recognition until you restarted Mira. It now switches as soon as you save.
Mira now asks the new models to keep their thinking short, so replies start sooner.
Gemini 2.5 Pro has been removed from the list, as it may not work with newly created API keys. When the AI declines to answer for safety reasons, Mira now says so instead of staying silent.
v2.2.0 updated the AI models to the GPT-5.6 / Claude 5 families and fixed missing responses with OpenAI. v2.1.x added microphone selection and input boost, and fixed an issue where API keys were not saved. v2.0.0 made Mira completely free.