Nexus STUDIO
Menu
Method Product Local-first Download

Desktop app · Pre-release

Build with AI.Without losing the thread.

Nexus turns an idea into work that was planned, tested and reviewed, instead of decisions buried somewhere in a chat log.

Download free See how it works
No account Your files stay on your machine Works with Claude Code
The Nexus Studio board, with work moving from backlog into planning, architecture and specs.

One board. One owner per stage.The real product interface

The method, in the open

Plan it. Build it. Prove it.

Every stage has an owner. You sign off on the result.

  1. 01

    Plan

  2. 02

    Architect

  3. 03

    Spec

  4. 04

    Build

  5. 05

    Test

  6. 06

    Review

Small changes take a shorter path. The route stays visible either way.

The product

Context that does not vanish into a chat.

Planning

Decisions live with the feature.

Brief, requirements, UX, architecture and epics sit together, readable by you and by the next agent that picks the work up.

  • Editable artifacts
  • Epics become issues
  • Existing projects
The planning screen, with product brief, PRD, UX, architecture and epics assigned to specialist agents.
Specialists

The right specialist for each job.

Nine sharply defined roles instead of one assistant trying to be all of them at once.

  • Live execution
  • Resumable sessions
  • Several agents at once
The Nexus Studio agents screen, showing the specialists and the model picked for each one.
Code knowledge

Agents find what already exists.

A navigable map gives you and the agents the same view of the repository.

  • Code relationships
  • Editable memory
  • Built-in terminal
The code knowledge screen, with repository statistics and a visual dependency map.

Local-first from day one

Your project stays yours.

Issues, plans and decisions are text files in your own repository. Turn the AI off and the tracker keeps working.

  • No Nexus cloud
  • No telemetry
  • No proprietary format
.nexus/issues/CORE-142.md
---
key: CORE-142
title: Electronic sign-off for timesheets
column: specs
origin: full
owner: agent:amelia
---

## Context

Readable by people, by agents and by git.
10 specialistsSharply defined roles.
38 skillsThe method ships with it.
0 accountsJust open a folder.
Plain textVisible in git.

Download

Install it and open a folder.

Free while it is in this phase. Builds for macOS, Windows and Linux, with no account, no sign-up and no cloud.

Choose your platform

All versions and release notes

The builds are not signed yet. Here is what to expect:

macOS will say the app is "damaged". It is not: an unsigned app carries no signature to check, and that is the wording Apple chose. Right-click and Open is not offered for this message. Drag the app to Applications, then run this once:

xattr -dr com.apple.quarantine "/Applications/Nexus Studio.app"

Windows shows a SmartScreen warning instead: choose More info, then Run anyway. Signing and notarisation are what remove both warnings.

Changelog

Every version, and what it contains.

Written by hand when a release is cut, and read by the installer, the app and this page alike. The entries are in English everywhere.

0.0.9

Added
  • Failed work offers a direct handoff to another installed engine. The new vendor receives the issue, specs, branch facts, tests and a bounded transcript of the recent conversation, while the original run remains in History.
  • A project-wide engine control can put every stage and agent on Codex or Claude Code at once. Per-stage and per-agent choices remain available for mixed projects.
  • Sprints can run continuously either as far as Review or through the full merge cycle, with a configurable concurrency limit, exact issue scope, a bounded decision ledger and Bob visible as the exception and completion steward.
  • A sprint can be removed with confirmation. Its file goes to Trash, its issues and specs stay, and membership is restored if the cleanup cannot finish.
Fixed
  • Codex runs now remain alive in the board, start and resume in the issue checkout, ask through the Studio question channel, receive the requested sandbox, show to-do progress and carry receipt timestamps on every normalised line.
  • Codex token totals now separate fresh, cached and output tokens without counting a resumed session twice. An absent monetary cost remains absent rather than becoming zero.
  • Changing engine no longer tries to resume a vendor-specific session id. Both explicit failover and an engine changed in Settings start a new vendor session with the persisted context.
  • A pending question blocks the board in the same tick that ends the run, so React state cannot advance a card past an unanswered decision.
  • A previous failed or finished turn is labelled as history while a later turn is working, instead of contradicting the live status at the top of the conversation.
  • Blocked issues cannot be dragged into code, and the board names the exact unfinished work that must move first.
Known
  • The native memory binary is not yet inside the installer. A machine without it installed falls back to Docker.
  • The native memory server does not yet survive a reboot on its own; the app starts it again when it is needed.

0.0.8

Added
  • A machine default for engine and model, set once in Settings and used by every agent that has no choice of its own. The Agents screen says where each choice came from and offers "use default" to clear one.
  • The dashboard's execution card now shows calendar time (how long at least one run was alive), how many runs ran in parallel on average, and time spent waiting on you. Each line says how many of the runs it was read from; nothing is estimated.
  • The quick path ends with the exit as a button: merge, pull request, or hand-off, naming the commit first when the lane has uncommitted work.
Fixed
  • The app killed the engine at the end of each turn, so an agent that had delegated to a subagent or scheduled its own wake-up was cut off, and the screen said finished over work still running. The run now stays open until the turn that owes nothing, or until the process exits; a subagent the process left behind reads "did not report" instead of done.
  • The next-step banner claimed "It wrote" a spec while the card still read planned: the quick flow's draft was being counted. Only a file saved under its own key counts, and the draft is no longer reported as a broken spec either.
  • The board lost its scroll position whenever an agent wrote a file: a read landing mid-write looked like a broken project, so the board unmounted and remounted. A doubtful reading over an open board is now re-read before it is believed.
  • A long question pushed the ask modal past the top and bottom of the window. It is capped to the viewport; the question scrolls inside its own region and the options and answer line stay on screen.
  • A mock embedding its images was refused at 512 KiB. The limit is 8 MiB; the sandbox, not the size, is what keeps a mock safe.
Known
  • A session cut off by an earlier version while its subagents ran may answer every resume with an empty turn. The app does not yet detect that or start a fresh session on its own; the next version will.
  • Calendar time per issue (backlog to done) is not shown: issues carry no timestamps for column changes yet.
  • Codex records carry no engine timestamps, so Codex questions do not count toward time waiting on you.
  • The native memory binary is not yet inside the installer, and the native server does not survive a reboot on its own.
All versions on GitHub

Pre-release, no spin

The product works.
Signing is not finished.

Pre-release. Every screen above is the product running.

Board and planning ready Live agents ready Automatic updates ready Builds for macOS, Linux and Windows Signing in progress
What works today?

Board, issues, specs, planning, sprints, documentation, configurable agents, live execution, questions asked mid-run, memory, terminal and editor. The schema battery runs on every build, and a good part of it is files broken on purpose that have to be refused: a check that cannot fail proves nothing.

What is still missing?

Public builds still need Apple notarisation and Windows signing. Git and pull requests inside the app were deliberately left out of the first version.

What do I need installed?

Claude Code. On launch, Nexus checks the other local requirements and tells you what is missing.

What the method rests on

EVO Method came first, not with the product.

EVO Method is an open method for building software with AI, forked from BMad Method and adapted independently. It exists on its own, outside this product, and Nexus Studio is the app that makes following it a matter of opening a folder.

Read EVO Method on GitHub

A public method, kept separately. Studio implements it rather than replacing it.

Built on

Other people's work.

Nexus does not do all of this by itself. Three independent projects are part of what you use here, and they deserve to be named.

  • Claude Code Anthropic

    The engine underneath. Every agent in Studio is Claude Code, running on your machine under your own subscription.

    Required
  • ai-memory Fabio Akita

    Keeps what the agents learn, per project and across projects, so the next session does not start from nothing.

    Optional
  • CodeGraph @tienmanh/codegraph, MIT

    Indexes the project so an agent finds what already exists instead of writing it again. The code map is drawn from it.

    Optional

Independent projects, credited for what they do. None of them sponsors, endorses or is affiliated with Nexus Studio or Evolution.

Pre-release

Keep the speed.
Lose the mess.

Download the build for your platform, or talk to us before you start.

Talk to us Community and support

The portal is where the community and support live. The email opens your mail client with the subject filled in, for reaching us directly.