Skip to content

AdultCoder · version 0.2

Pro-grade AI coding. Your data never leaves your control.

Building a whole ecosystem of enterprise AI coding tools: simple to use, with privacy and compliance designed in, for software where both are critical.

A terminal shell and a Mac app. Beta soon.

Join the waitlist

“Built for privacy-first teams and enterprises. Not for kids.”

Proof, not promises

It built most of this site. And my personal one.

The lines here are real, from those sessions: it planned, wrote and checked the work, compacted its own context when the conversation grew long, and committed as it went.

Visit my personal site nitinjaiswal.com

AdultCoder
  1. todo_writeTask list saved: 0 of 8 done. In progress: Scaffold Astro project + install.
  2. write_filewrote 3569 bytes to src/components/home/Proof.astro
  3. edit_fileapplied 2 edit operation(s) to src/components/home/Hero.astro
  4. bash[build] 39 page(s) built in 753ms
  5. Compacted the conversation: 82 earlier messages, about 217K tokens, are now a summary, and the work goes on from it.
  6. Tasks
    • [x] Scaffold Astro project + install
    • [x] Subtle sound: opt-in toggle, ambient pad, UI sounds
    • [x] Forbidden-string scan script over src + dist, fix hits
    • [x] Final QA: Chrome + WebKit, both themes, 390 and 1440
  7. git_commit[main d86756f] Keep the hero's notes on screen and give the glass its blur back
Real lines from the sessions: AdultCoder built most of this site, and my personal one.
AdultCoder at work in a terminal on my personal site: file reads, writes and checks scroll past, the conversation is compacted, and the task list shows five of eleven steps done
The terminal, mid-session on my personal site · the folder names are blurred

Plans first. Sandboxed. With brakes.

Every run

Six things it does, so the work arrives finished and accounted for.

Plans first

It plans the work, and builds once you approve.

Sandboxed

Every command runs in a sandbox.

Brakes

Auto mode, with brakes: it stops before anything destructive.

Your keys stay yours

The agent never sees them.

Receipts

An audit log that can’t be quietly edited: one shared record.

Budgets

Set the budget. Get the finished work.

Your models, your hardware

The model is half the result. The tooling is the other half.

  • On your own servers

    Open-weight models, running on hardware you control.

  • Through your own keys

    Hosted models, through your own keys, when you choose them.

Long runs

Built to finish, not just to start.

Compacted the conversation: 82 earlier messages, about 217K tokens, are now a summary, and the work goes on from it.From a real session.

  • It compacts long workWhen a conversation grows long, it folds the past into a summary and carries on.
  • It waits out outagesA provider goes down; the session waits, then resumes where it was.
  • It changes modelsA session can be resumed on another model, replayed, undone, or shared.

A native engine, not a web app in disguise.

To start
0.01 s
Peak memory for a turn
56 MB
A full turn in a Rust repository of
200,000 lines

A native engine, not a web app in disguise · measured for adultcoder.tech

Lines of Rust, about
250,000
Tests, about
2,400
Since
March 2026

Native, next

A native desktop app is on its way.

Its UI is being proven first, in a separate proof of concept: a native agent chat, written in Rust on my own UI toolkit and drawn on the GPU.

See it in the story

A native agent chat in light mode: a conversation titled “Port the thumbnail worker from Python to Rust”, with highlighted Python code, two replies and a bar chart, beside a list of chats
The same native agent chat in dark mode, further into the conversation: a follow-up question, a reply with the chart, and the next answer arriving

Build more. Worry less.

Built for privacy-first teams and enterprises. Join the waitlist for the beta.

Ideas by me, built with AdultCoder (opens adultcoder.tech in a new tab)