I used to open X to *reply * to a few people.
An hour later I was still rewriting drafts.
Not because I had nothing to say. Because every reply took too long. Read the post. Find a take. Soften it. Make it sound like me. Delete. Start over.
That friction is annoying if you are building in public and actually want to talk to people. So I built XReply.
What XReply does
XReply is a Chrome extension for X (Twitter).
- Click a post
- Get an AI reply in your voice
- Paste it
About 2 seconds. Not generic “Great point!” energy. Something closer to what you would have written if you were not tired.
Free to start with 200 credits. Then $15 for 5,000 credits. No subscription.
Site: https://xreply.net
The problem I kept hitting
Most AI reply tools fail in the same way:
- They sound like ChatGPT
- They ignore your tone
- They push you into another tab or dashboard
I did not want another writing app. I wanted something that stays inside the reply flow on X.
What I built
Stack, kept simple:
- Chrome extension for click → generate → paste
- Next.js web app for auth, profile, billing, dashboard
- Supabase for users and credits
- Paddle for one-time credit purchases
- An LLM API for reply generation
Core idea:
- Learn the user’s voice from their profile (tone, bio, interests, what to avoid)
- Generate a short reply that matches that voice
- Charge 1 credit per reply
- No monthly plan. People top up when they run out.
Product decisions that mattered
1. Voice over cleverness
A funny reply that does not sound like you is useless. Matching tone was the main goal.
2. Stay in the workflow
No extra dashboard for the act of replying. Click the post, get the reply, paste.
3. Credits instead of subscriptions
Some people reply a lot. Some barely reply. Prepaid credits felt fairer than locking everyone into a monthly bill.
4. Ship thin, then improve
First version: profile + generate + paste. Everything else came after real usage.
*What surprised me *
People care less about “AI” and more about one thing:
> Does this sound like me?
When it does, they keep using it. When it does not, they bounce fast. So most of the work went into prompting, profile context, and keeping replies short enough for X.
What’s next
- Better tone matching from real usage
- Faster generation
- More polish in the extension UX
If you reply a lot on X and want something that sounds like you, try it here:
https://xreply.net
I am also happy to answer questions about the extension, the credit model, or shipping a small AI product as a solo founder.
Top comments (2)
Super interesting project! I'm curious how you technically handle the "in your own voice" aspect.
Good question! It's a one-time profile setup — you fill in your role, background, interests, and goals once, and that gets fed into the prompt every time a reply gets generated.
The part that actually matters isn't the profile itself though, it's the rule sitting on top of it: I explicitly tell the model to only pull in your background when the post genuinely connects to it, and to reply naturally otherwise. Without that constraint, every reply turns into a mini elevator pitch about yourself, which is exactly the "sounds like AI" problem I was trying to avoid.
There's also per-user tone/length settings, and I recently tuned the sampling so regenerating on the same post doesn't just spit out the same phrasing twice. Happy to go deeper on any part of it if you're curious!