About AlbertCode
The bottleneck was never writing the code.
Engineering organisations spent decades building the apparatus that assigns responsibility for a change — review, testing, sign-off, audit. An agent that skips all of it isn't faster; it has moved the risk somewhere nobody is watching. AlbertCode keeps that apparatus and puts the agent inside it.
Mission
From idea to working software — with the proof attached.
AlbertCode exists to make AI software engineering something a team can adopt and defend: every change planned before it is written, built away from your working tree, verified by your own tests, and accepted by a person who can see exactly what they are accepting.
The model does the reasoning. AlbertCode decides what it may touch, in what order, under whose approval, and what it leaves behind. The goal is AI-assisted engineering as ordinary infrastructure — as dependable as a build pipeline, and as unremarkable.
Built for developers everywhere. AlbertCode is an AI software engineering initiative with global aspirations: AlbertCode SWE Agent today, installable on Windows, macOS and Linux, with no account and no fee to start.
Principles
How the work is judged.
01
Evidence over assertion
A change should arrive with its proof: the plan that was approved, the commands that ran, the checks that passed, and the exact diff. Claims about the product follow the same rule — a number is published only with its evidence.
02
People hold the authority
An agent can draft a plausible change to almost any codebase. It cannot carry responsibility for it. So the default is two human gates, and the fast lanes are choices you make out loud.
03
Private by default
The agents run on your machine with no account and no telemetry. Your code goes to the model you choose — or nowhere, with a local model.
04
Honest boundaries
Say what is built, and what isn't. Label illustrations, keep self-reported results apart from verified ones, and write “not documented” rather than guess.
Origin
It began as research into governed autonomy.
AlbertCode was conceptualised, architected, engineered and developed by Huy Nguyen, as a study of how an autonomous coding agent can be bounded by authored policy, held to human approval, and made to leave evidence that survives scrutiny.
- Can the boundary of an agent's authority be declared as code and enforced before it acts?
- What evidence must a change carry for a person to be accountable for approving it?
- Which controls survive an agent that is capable, mistaken, or adversarially steered?
Where it stands
What we don't claim.
Stating the boundary plainly is part of the point.
- The installers aren't code-signed yet, so macOS and Windows warn the first time you run a downloaded one.
- Linux support is x86-64 with glibc only: no ARM, no Alpine or other musl systems.
- The published benchmark is self-reported and hasn't been independently verified.
- The security and quality reviews are made by the model you connect: they record that a review happened, not that it was right. Human review remains required.
- Results depend on the model, the request and the quality of the repository's own tests.
- Neither agent is offered as a hosted service; both run only on your computer.