You describe the outcome. Nexus plans, builds, verifies and repairs it — in your repos, your documents, your research. Your keys, your models, your data. All local.
It doesn’t chat.
It finishes.
▸ step 02 — surfaces
Three tabs. Your agent context, providers, browser and skills follow you everywhere.
The agent that finishes things. Chat, research the web, and get real files — Word reports, decks, spreadsheets, PDFs, LaTeX — and inline interactive charts validated and delivered to your Nexus folder.
docx · pptx · xlsx · pdf · charts · tex
Plan / Ask / Auto control over any repo. It inspects, plans, implements — then a verify supervisor checks the diff, runs test cascades, and repairs what fails before declaring done.
LSP enabled · repo map · subagent fleet · plugins
Q&A over your PDFs, Office docs, YouTube videos and web pages — every claim cited, Net-Guard SSRF protected, and refuses to guess when your files don’t cover it.
hybrid retrieval · studio exports · net-guard safe
mode 01 — home
Home isn’t a chatbot with ambitions — it’s a production line. Say what you need; the research, drafting, charting, and polishing happen in the background, and a validated file lands in your Nexus folder.
mode 02 — code
Code mode pairs a long-horizon agent with a verify supervisor. It plans before touching anything, implements with real tools, then checks its own diff — and repairs what fails — before you ever see “done”.
mode 03 — notebook
Feed it your PDFs, slides, docs, YouTube videos and web pages. Every claim carries a clickable citation to the exact passage and a groundedness score. If your sources don’t cover it, Nexus refuses rather than guessing.
Every run ends
in proof.
▸ step 03 — how it thinks
LangChain tools and models · LangGraph workflow and checkpoints · DeepAgents long-horizon loop. Watch the run:
| mode | what the agent may do | recursion budget | auto-repairs |
|---|---|---|---|
| Plan | Read-only — no writes, no shell, no MCP | 40 | 0 |
| Ask | Interactive — proposes, you approve | 100 | 3 |
| Auto | End-to-end autonomous with verification | 150 | 5 |
Rate limits retry with bounded backoff and resumable checkpoints. Smart per-target loop prevention redirects the agent from unhelpful web queries. A doom-loop breaker stops stuck tool calls. Every run ends in a checkpoint you can “continue” without repeating completed steps.
Your keys. Your machine.
Your rules.
▸ step 04 — trust
Risky commands open a modal: Deny · Allow once · Allow for session. Silence for 120 s means deny.
~25 destructive patterns are always refused — disk wipes, fork bombs, registry deletes, boot tampering.
Plan mode hard-refuses writes, edits, shell and MCP. Explore a codebase with zero side effects.
Keys encrypted with the OS keychain; terminals get scrubbed env; daemons run on a strict allowlist.
Your code goes only to the provider you configure. State is a local file; transcripts are local JSONL.
.nexus/permissions.json allow/ask/deny globs, plus Net-Guard SSRF security blocking private IP access.
The honest fine print: Nexus runs real shell commands and edits real files — that’s the point. Its safety model is approvals + deny-backstop + encryption, not a sandbox. Review the Diff tab before keeping changes.
bring your own key
frequently asked
Yes — Nexus is bring-your-own-key across 15 providers. Add them in Settings → AI Providers (encrypted at rest) or via environment variables. Prefer fully local? Point Nexus at Ollama and the offline embeddings fallback — no cloud calls at all.
Only to the LLM provider you configure — there is no Nexus cloud, no telemetry service, no middleman. App state lives in one local JSON file; transcripts are local JSONL; secrets are encrypted by your OS keychain.
Autocomplete tools suggest lines as you type. Nexus is an autonomous agent for your whole desktop: it takes a goal, plans the implementation, edits files and executes terminal commands, verifies its own work against test suites, repairs failures, provides per-hunk undo, and extends beyond code to document deliverables and grounded research in the same app.
It can make changes — that’s the point — but you’re guarded: approval gates on risky commands, a hardcoded deny backstop for destructive patterns, read-only Plan mode, isolated git worktrees, 20-deep checkpoints, per-hunk discard, and Nexus-only commit resets that never touch your own commits.
Every Notebook answer is scored 0–10 against the evidence retrieved from your sources. If your files don’t cover the question, Nexus tells you “not in your files” and refuses rather than guessing. Citations open the exact source passage.
Windows 10 and 11 (64-bit) — available as an auto-updating NSIS Setup or a standalone Portable exe. You can also build from source anywhere Node.js 20+ runs.
The Windows Setup ships with electron-updater: a pill in the app shows download progress and restarts to apply. Portable builds re-download. New versions land on GitHub Releases — this page always links the latest.
Nexus includes an in-app extension store in Settings where you can discover, install, and manage plugins. Plugins bundle skills, specialist subagents, coding rules, slash commands, and lifecycle hooks into self-contained units. You can install official plugins, connect custom registry feeds, or link local developer folders to test plugins live.
Extensively: browse the in-app Plugin Marketplace in Settings to install pre-packaged engineering and research suites with one click. You can also drop SKILL.md files into .nexus/skills, define custom slash commands, add project rules (AGENTS.md, CLAUDE.md, .cursorrules…), connect MCP servers (stdio/HTTP/SSE), and set allow/ask/deny permission policies per project.
▸ step 05 — get it
Links resolve to the latest GitHub release — currently v…. The Setup auto-updates; the Portable exe re-downloads.
Auto-updating installer or a portable exe — no install, just run.
git clone https://github.com/amine-elhanine/nexus-code.git
cd nexus-code
npm install
npm run dev # develop
npm run dist # package installers to release/Requires Node.js 20+ and npm.