███╗   ██╗███████╗██╗   ██╗██████╗  █████╗ ██╗
 ████╗  ██║██╔════╝██║   ██║██╔══██╗██╔══██╗██║
 ██╔██╗ ██║█████╗  ██║   ██║██████╔╝███████║██║
 ██║╚██╗██║██╔══╝  ██║   ██║██╔══██╗██╔══██║██║
 ██║ ╚████║███████╗╚██████╔╝██║  ██║██║  ██║███████╗
 ╚═╝  ╚═══╝╚══════╝ ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝
       ╦╦ ╦╔╗╔╦╔═╦╔═╗
       ║║ ║║║║╠╩╗║║╣
      ╚╝╚═╝╝╚╝╩ ╩╩╚═╝

Neural Junkie

Talk to any AI model. Local or cloud. One workspace.

Open-source multi-agent desktop hub — bring your own model (Ollama, Claude, OpenAI-compatible, LM Studio, Cursor/Gemini CLI), run specialist agents from domain packs, and keep humans in the loop with file-change and tool approval. Local-first by default.

Any model	Specialist agents	Human control
Per-agent providers, local or cloud	Packs for software, IDE, AWS, life sciences, and more	Approval gates, scoped context — not a shared hive mind

Agents, tools, and runbooks are portable composition units you can share and hydrate — not hub-locked state. See the Composition Model and Context Stack.

Site: camronwood.github.io/neural-junkie · 5-minute guide: start-here.html

Install

macOS (Homebrew):

brew tap camronwood/tap
brew install --cask neural-junkie

Linux (Homebrew):

brew tap camronwood/tap
brew install neural-junkie

See docs/HOMEBREW.md for upgrades and tap details.

All platforms — GitHub Releases: latest release or the downloads page:

Platform	File
macOS (Apple Silicon)	.dmg with aarch64 in the name
macOS (Intel)	.dmg with x64 or x86_64 in the name
Windows	.msi installer (or .exe setup)
Linux	.deb (x86_64)

The Go hub is bundled as a Tauri sidecar — you do not need Go installed to run the desktop app. Use the Tauri desktop app for the full workspace; http://localhost:18765 is a lightweight browser chat client only.

macOS: Official Release builds are ad-hoc signed. If Gatekeeper blocks first launch, right-click → Open. Notarization is planned for a later stable cut — see docs/INSTALL_TRUST.md.

Ollama: macOS installers bundle the Ollama runtime; Windows/Linux use slim installers and the setup wizard can install Ollama on first launch. First run pulls a default model once (internet required that one time). Cloud APIs remain optional.

Auto-update: macOS and Windows check for updates in-app (Settings → About → Check for updates). Linux .deb upgrades are manual. Details: docs/RELEASE_UPDATES.md.

Quick start after install: docs/DOWNLOAD.md

Five-minute first win

Fresh installs start lean: Assistant plus any auto-detected CLI agents (Cursor, Gemini, …). Domain packs are off until you enable them.

1. Open Neural Junkie and finish the setup wizard (local Ollama and/or a cloud key).
2. Settings → Domain packs → Pack store — install and enable Software development (specialists) and optionally IDE (editor depth).
3. Create a repo expert: /create-repo-agent /path/to/repo MyRepoExpert
4. Ask: @MyRepoExpert summarize the architecture and top risk areas
5. Collaborate: /collaborate @BackendEngineer @SecurityReviewer harden auth middleware — review proposals and approve file changes when prompted.

More: docs/USER_VALUE_GUIDE.md · marketing start-here

What you get

• BYOM — Bundled Ollama, Claude, OpenAI-compatible APIs, LM Studio, and CLI agents; switch providers per agent or globally
• Multi-agent workspace — Channels, DMs, threads, bounded /collaborate sessions, Slack Connect, human approval for file edits and tools
• Neural Canvas — Durable agent reports, charts, timelines, diagrams, and workbench artifacts beside chat
• IDE v4 + NJ Fix Loop — Monaco LSP, Ask/Agent composer, remote SSH via nj-remote, closed-loop boot/build repair (IDE pack, IDE v4)
• Domain packs (12) — IDE, software development, life sciences, CAD, specialist tuning, AWS, incident management, web browser, music creation, model arena, maps, room-chat — plus Pack Dev Studio for custom packs (PACKS.md)
• Composition + scoped context — Share Agent bundles, MCP tool grants by name, runbook export/import; conversation context is stacked and scoped, not a shared brain
• Repo + Confluence experts — Index codebases and doc spaces; MCP export/import for sharing agent knowledge

Full guides: docs/features/ on the marketing site.

Screenshot

Main workspace — files, editor, multi-agent collaboration, and chat (same as the marketing site).

Agents (after packs)

On by default	How you get more
Assistant — reminders, tasks, notes, chat guidance	Enable Software development for BackendEngineer, FrontendEngineer, PlatformEngineer, SecurityReviewer, SoftwareArchitect, CodeReviewer
CLI agents (Cursor, Gemini, …) when on PATH	Enable IDE for editor/LSP depth; create repo/Confluence/expert agents with slash commands

Type / or click / for the command palette. /help lists everything. Details: docs/SOFTWARE_DEVELOPMENT_PACK.md, docs/CLI_AGENTS.md, docs/REPO_AGENTS.md.

Quick start (from source)

git clone https://github.com/camronwood/neural-junkie.git
cd neural-junkie

make gui-install   # first time: desktop deps
make start-all     # hub (in-process agents) + desktop app

make start-all does not spawn separate cmd/agent processes. Optional splits:

make server    # hub only
make gui       # desktop only
make chat      # terminal chat
make agents    # six specialists as OS processes (avoid duplicate names vs in-process config)

Configure providers in the Setup Wizard or Settings → AI Providers. For source builds without a bundled runtime:

make pull-models   # Ollama: qwen3.5:27b + qwen3.5:9b (or install from https://ollama.ai)

Cloud keys: Settings UI, or cp env.example env.local and set ANTHROPIC_API_KEY. Slash switches: /switch-provider, /switch-all-providers. Full setup: docs/GETTING_STARTED.md.

Project structure

neural-junkie/
├── cmd/                 # server, agent, chat, cli, nj-remote
├── desktop/             # Tauri + React app (src + src-tauri)
├── internal/            # hub, agents, AI providers, packs, repo, MCP, …
├── packs/               # official catalog (catalog.json)
├── scenarios/           # release / user-flow scenario fixtures
├── assets/              # screenshots, marketing images
├── docs/                # documentation + marketing site HTML
├── scripts/             # build and release
└── .github/workflows/   # CI / release

Documentation

Full index: DOCS.md (make docs prints the same file).

Doc	What it covers
Getting Started	Source setup, configuration, make targets
Domain packs	Official packs and install flow
Composition Model	Portable agents, tools, runbooks
Context Model	Scoped conversation context stack
Implementation sessions	IDE Agent mode, verify/repair, NJ Fix Loop
IDE v4	Full LSP, remote SSH, dev containers
Collaboration	Multi-agent planning, delegation, execution
Architecture	System design, data flow, patterns
Changelog	Version history
Known issues	Beta limitations and workarounds
Contributing	How to contribute and run test gates

Make targets

make start-all        # Hub + desktop
make server / gui / chat / agents / stop / refresh
make gui-install      # Desktop dependencies
make build            # Go binaries
make test-go          # Go tests (-count=1)
make test-all         # vet + Go + desktop tsc + Vitest
make pull-models      # Ollama models
make bundle           # Current-platform distributable
make release VERSION=…  # Bump, commit, tag (then push for CI)

See the Makefile and docs/GETTING_STARTED.md for the full list.

Built for $0

Neural Junkie is an experiment: free for users and built, shipped, and marketed with free open-source tools only. See docs/BUILT_FOR_ZERO.md, docs/THANKS.md, and built-for-zero.html. Install trust notes: INSTALL_TRUST.md.

License

MIT

Contributing · Security · Issues

• Contributing
• Security policy
• Thank you — open source
• Known issues
• Report a bug