Nexus AI

A local workshop for your AI coding agents.

Run Claude Code, Codex, OpenCode, Hermes (and soon Gemini CLI, Qwen Code) in a real IDE, one at a time or as a team:
a conductor agent plans the work, the others build it in parallel in an isolated copy, talk to each other while they
code, get reviewed, and you merge the result.

Download · Features · Build from source ·
Contributing · User guide (FR) · Architecture

_{Orchestra: one goal, one conductor, a team of agents that talk to each other.}

[!NOTE]
Status: preview 0.2.1 — Linux is the verified platform. Windows and macOS builds are produced but experimental:
no sandbox, and process/terminal behaviour has not been verified on real machines. Details:
.coordination/STATUS.md · release notes.

The key principle

The agent CLIs do the coding — Nexus never replaces them. Claude Code, Codex, Gemini CLI, Qwen Code, OpenCode and
Hermes keep their own models, tools, logins and subscriptions; they read and edit files with their own tools. Nexus
launches them, isolates them (worktree copy, optional sandbox), orchestrates them (plan, task graph, hand-offs),
relays messages between them, reviews their results and merges them into your project when you confirm.

Screenshots

Compare agents — same instruction, each agent in its own isolated copy	IDE workspace — explorer, Monaco editor, real terminals, Git
	Agent detection — Nexus finds the CLIs installed on your machine, their version and whether they are signed in. No API key to paste: each agent keeps its own login and subscription. Screenshots show the French UI; English is available and follows your OS language.

Download

Builds are attached to the Releases page:

SHA256SUMS.txt lists the checksum of every file. Install and sign in to at least one agent CLI first; Nexus detects it
automatically (the onboarding screen shows the install command of each one).

Agents

Features

• Sessions — one agent on one task, in its real terminal (interactive TUI) or as a conversation. Three modes:
  • Supervised (default): the agent works in an isolated Git worktree; you review the diff and accept or ask for changes.
  • Autonomous: the agent works directly in your project folder.
  • Sandboxed (opt-in, Linux): isolated copy + bubblewrap, seccomp, no direct network (proxy limited to the agent's
    provider). The CLI's own login folder (e.g. ~/.claude) is mounted read-write so it can work; .env files are
    visible to the agent (per-project protection exists in the engine but has no switch in the UI yet).
• Broadcast — send the same instruction to several agents at once, each in its own copy, and compare.
• Orchestra — a conductor agent (read-only) plans the goal as a dependency graph; tasks run in parallel in a shared
  isolated copy; agents talk through the nexus MCP team bus (team_status, send_message, read_messages,
  ask_lead); the conductor reviews each result; you merge in one confirmed click. Survives an app restart (resume),
  can be stopped per mission, reassigns tasks of a failing agent.
• IDE workspace — file explorer, Monaco editor, PTY terminals, command palette, VS Code-like shortcuts, Git
  (source control, diff), pull requests via gh.
• Deployment planner — proposes hosting setups (Vercel, Render, Railway, Hostinger VPS, MongoDB Atlas) and runs them
  step by step, each step confirmed.
• Usage view — quotas and usage per agent, never invented (shown as unavailable when unknown).
• Light / dark / system themes; French and English (first-run language follows the OS).

What is verified only against fakes: GitHub PR creation, deployment providers. Voice dictation needs whisper.cpp or
explicit cloud consent and has never been run for real.

Security in short

Untrusted renderer (no Node, strict CSP in production), Zod-validated IPC on both sides, no shell: true, OS keychain for
secrets, append-only local SQLite event log. Each agent runs inside a process boundary (Linux: PID namespace + parent-death
kill; macOS: process group; Windows: taskkill tree — the last two unverified). Git metadata is read-only inside the
sandbox and host Git refuses repository configs that define commands. See the threat model.

Build from source

Requirements: Node 24.18, pnpm 11 (never npm), Linux with bwrap, socat and unprivileged user namespaces for the
sandbox tests.

pnpm install --frozen-lockfile && pnpm run generate
pnpm run check                          # lint, format, typecheck, tests
pnpm start                              # run the app in development
pnpm --filter @nexus/desktop run make   # build packages into apps/desktop/out/make

Reproducible Linux build in Docker: docker build -f Dockerfile.release --target export -o dist .

Contributing

Nexus AI is young and there is a lot of room to make a real impact. Contributions of every size are welcome — and you
do not need to write code to help:

• Try it and report — run it with your agent CLIs and open an issue
  for what breaks or feels wrong.
• Verify a platform — Windows and macOS builds have never been confirmed on real machines; a test report is
  extremely valuable.
• Run an untested agent — Gemini CLI and Qwen Code adapters exist but have never been run for real.
• Translate — the UI ships in French and English; strings live in apps/desktop/src/renderer/i18n/.
• Pick a good first issue
  or one labelled help wanted.

Start with CONTRIBUTING.md. Please follow our Code of Conduct; report
vulnerabilities privately as described in SECURITY.md. Accessibility commitment and known limits:
ACCESSIBILITY.md. Publishing: docs/RELEASING.md · guide for coding
agents: CLAUDE.md.

If Nexus AI is useful to you, a ⭐ on GitHub helps other developers find it.

License and notices

Nexus AI is licensed under the Apache License 2.0. See NOTICE for project and third-party notices. Copyright 2026 Dayifour.