Mitra is not a chatbot. There is no streak, no progress bar, no badge to earn. The whole point is that you leave a little more grounded than you arrived, and then you leave.
“Does this bring you one step closer to your own wisdom — or does it make the app more indispensable than the practice?”
Mitra is built to answer the first question with yes, and refuse the second. That's why the call ends on purpose. That's why there's no streak.
Different needs use different tools. Pick the one that matches the moment, not the one with the most features.
Three taps — mood, energy, and an optional one-liner about what's going on — give you a two-to-five minute audio practice. Then you close your eyes and do it.
A short voice call with Mitra. You speak, she speaks. Up to ten exchanges. She listens more than she answers. Replies are one or two sentences. The call ends on purpose.
Eight tiles — anxious, sad, tired, okay, good, overwhelmed, numb, or angry. The truest one, not the most flattering one.
Low, medium, or high — how the body actually is, not how you wish it were.
One line about what's going on. Mitra uses it to choose a practice. Skipping is equally fine — many of the best practices need no context.
Bill's voice reads a 2 to 5 minute practice. There's a 15-second skip-back. You can show the words if you'd rather read along. When you're done, you tap “I'm done.”
Your browser will ask for microphone permission once. Allow it for this site.
Bill's voice: “How are you arriving today?” Then it's your turn — speak naturally, no need to press anything.
There's a visible counter — turn N of 10. Replies are short by design. Pauses are part of the practice; don't rush to fill them.
Behind the scenes, each thing you say is checked for real distress signals before the call continues. If something serious comes up, the call ends gently and you're shown a guide-coming card with hotlines. You don't have to know it's happening — it just is.
When you hit ten turns, Mitra closes the call with “That was enough. Come back tomorrow if you'd like.” You can also tap “end the call” any time.
Both doors have the same safety net underneath. Every thing you type or say is checked for real distress signals — self-harm intent, abuse, immediate danger — before Mitra responds. If something serious comes up, the app stops what it was doing.
Mitra serves a short pre-vetted grounding practice — not an LLM-generated one — and shows the guide-coming card alongside it. You aren't alone in the minute it takes for help to reach you.
Mitra disconnects the call and you're taken to the same guide-coming card. The call doesn't spiral; a human is the next step, not more conversation with an app.
If you are in immediate danger, please use a hotline directly — they are staffed right now and can do more than an app.
Not yet. Bill, the current voice, is a stand-in until Nitesh records his own voice and we plug it in.
Not in Phase 1. Every call starts fresh. The reason: cross-session memory needs a real account, which we haven't built yet — and we'd rather have no memory than half-built memory.
Claude Opus 4.7 for the conversation and practice generation. Claude Haiku 4.5 for the safety classifier. We picked Claude over GPT-5.5 after a 24-sample bake-off — same voice quality, perfect adherence to the duration rules.
Because the most likely failure mode for any AI companion is that it becomes a place to spiral. Ten turns is enough to be heard, not enough to substitute for a friend or a teacher.
That's a sign you need a human, not an app. The hotlines above are staffed, and if you want a real teacher from mteen, the crisis card connects you to one.
mteen — Nitesh Batra and team. The teaching corpus is from Nitesh's 20 years of contemplative work (Ashtanga yoga, Compassion Cultivation from Stanford CCARE, Lojong, Bhagavad Gita). The voice and software are stand-ins for the human teaching, not replacements.