Any model
Talk to local Ollama (bundled on macOS; wizard on Windows/Linux) or cloud APIs — Claude, OpenAI-compatible, Gemini, and more. Hybrid routing: iterate cheap locally, send hard reviews to the cloud.
Open source · Any model · Local or cloud · v1.2.0-beta.22
Talk to any AI model.
Local or cloud. One workspace.
Neural Junkie is an open-source desktop app where you chat with any model you choose — bundled Ollama on your machine, or cloud backends like Claude, OpenAI-compatible APIs, and more. Then run them as a team of specialist agents — backend, security, biology, AWS, or experts you create — with human approval and optional IDE v4 when you need an editor.
Download for macOS / Windows / Linux
Homebrew (macOS / Linux)
5-minute start guide
Built for $0
See use cases
-
Claude
-
GPT
-
Gemini
-
Ollama
-
HF
-
Cursor
-
LM Studio
your AI command center
Neural Junkie
Neural Junkie
Any model local Ollama or cloud APIs — pick per agent
Teams Slack Connect + threaded chat
You custom experts, learning, collaboration
Three pillars
Not locked to one vendor chat — run any model where you choose, then specialize, collaborate, and stay in control.
Specialist agents
Backend, security, biology, CAD, AWS — or /create-expert for any domain. @mention the right role instead of one model pretending to be everything.
Human control
Approve file changes with diff preview. Bounded /collaborate sessions. Tool gates. Personal learning only when you confirm.
Built for how you work
Same hive-mind workspace — different packs and experts for your job.
Software teams
Engineering specialists, repo agents, Git in the app, IDE v4 with full LSP and remote SSH. Ship with approval gates, not silent edits.
IDE v4 guide →Life sciences & lab
BiologyExpert, OpenBioLLM, sequence tools, and protein folding — local when you want it. Customer packs add instrument-specific QC without a public store listing.
Lab workflows →Platform & incidents
AWSExpert with read-only CLI MCP, IncidentManager for Jira triage, and handoff to engineering specialists when code changes are needed.
Domain packs →Any domain
Guitar coach, compliance reviewer, internal API expert — custom experts are first-class. Slack Connect keeps teams in sync when you need it.
Experts guide →See it in action
Short demos from the real desktop — product overview, multi-agent chat, Assistant workflows, Slack, local image gen, and away mode.
The workspace in 40 seconds
Feature fly-through — specialists, collaboration, IDE, packs, and local AI in one desktop app.
Gemini, Copilot, Cursor — one channel
Gemini, Copilot, and Cursor agents in the same room — the multi-agent pattern you use for real work.
Ask the Assistant
Your always-on Assistant — reminders, tasks, workspace context, and command help in one DM.
Slack Connect
Mirror bound channels, route @mentions to any agent, and keep your team in Slack while agents run locally.
Local image gen for free
Generate images with local Ollama models — no cloud API required for creative workflows.
Agents respond when you're away
Slack away mode and personal inbox — your agents keep answering while you're offline.
Flagship capabilities ·
Multi-agent collaboration
Run structured agent projects, not endless @mention chains
/collaborate turns a goal into a bounded session: shared plan, review rounds, human approval, then execution with per-agent tasks and explicit workspace confirmation before agents touch your files. Great for research write-ups, cross-functional plans, and multi-step projects — not just code.
- Phases you control — planning → review → approve → execute with visible status.
- Runbooks v2 — persisted definitions, run history, connector profiles, and replay from the desktop library.
- Personal learning — collaboration-scoped memory when Specialist tuning is on.
- v1.2.0-beta.22 — Semantic stamp dual-gate graduation, continuity dialogue, and conversation-memory retrieval gates. Known issues.
1Plan
2Review
3Approve
4Execute
Install-and-go local AI
Download once. Run locally. Wizard handles Ollama.
macOS installers bundle the Ollama runtime; Linux and Windows use slim builds and the setup wizard Install Ollama on first launch. Complete the wizard, pull a default model once, and your agents run on your hardware — private by default, cloud optional. After first run, browse more models in the built-in Model library.
- Auto-start — production app runs
ollama servewith models in app data. - Linux & Windows — wizard installs Ollama (winget, silent installer, or
install.sh). - Model library — curated Ollama + Hugging Face grid (⇧⌘M) when you want more than the default.
- Cloud when you need it — Claude, OpenAI-compatible APIs, and per-agent routing in Settings.
Download installer
Pull model once
Multi-agent chat
Slack integration
Your team stays in Slack. Agents stay in sync.
Slack Connect mirrors bound channels into Neural Junkie with two-way threading: @mentions and channel lines route to the agent you choose — Assistant, custom experts, repo agents, or any configured Claude / GPT / Gemini / Ollama backend. Personal inbox: DM the NJ bot from Slack mobile; replies land in a private hub channel. Setup diagnostics in Settings help you verify OAuth, Socket Mode, and channel routing.
- Per-channel policy — always-on, questions-only, or @mention routing.
- Mobile → agent — DM the bot; pick which agent answers in NJ.
- Forward from channels — @mention,
nj:prefix, or reaction emoji rules. - Real names in NJ — Slack channels show as
#cursor-test, not opaque IDs.
#cursor-test
#cursor-test
Domain packs
Install packs for the work you do — engineering, research, design, cloud, LAN rooms, or tuning
Open Domain packs from the toolbar chip (⌘⇧K) → Store to install official packs from GitHub. The hub merges an embedded catalog snapshot so new packs appear without a restart. Enable multiple packs; pick which enabled pack controls IDE vs team layout. Chat, Slack, and custom experts stay available either way.
Software development
Engineering specialists, Git in the app, IDE v4 (full LSP, remote SSH, @codebase), implementation sessions.
Life sciences
BiologyExpert, OpenBioLLM, sequence tools, protein folding — plus customer sideload packs for instrument-specific QC.
CAD
CADExpert, OpenSCAD workbench, param sliders, Three.js STL preview — parametric design in chat.
Specialist tuning
LoRA training from chat and collabs, personal learning opt-in, adapter compose in Ollama.
AWS
AWSExpert with read-only AWS CLI MCP tools, SSO profile picker, and account-scoped infra queries.
Incident management
IncidentManager for Jira triage, reproduction steps, and handoff to engineering specialists.
Web browser
WebBrowserExpert, HTML preview workbench for .html files and local dev servers, plus fetch_url.
Music creation
MusicExpert, ACE-Step generation, music workbench, stems export, and EP workflow.
Model Arena
Head-to-head model matches on chess, Connect Four, and logic puzzles with verifiable scoring and a live Arena workbench.
Room chat
Same-room LAN chat — host ephemeral rooms, guests join with a code, and @mention the host’s agents. No cloud hub.
Custom packs: Pack Dev Studio + sideload zip for customer bundles. See pack catalog and PACKS.md.
IDE v4 · Software development pack
When you need an editor, build the one you own
IDE v4 is a capability on top of the hive-mind — not the whole product. Full Monaco LSP on local workspaces, remote SSH via nj-remote, dev container attach, and tree-sitter symbols — with Ask / Agent composer and implementation sessions from IDE v3 underneath.
- Full LSP — diagnostics, hover, completion, go-to-definition, rename via
gopls,rust-analyzer,pyright-langserver. - Remote SSH — EC2, dev box, or container; file tree, git, terminal, and
@codebasethrough one backend. - Open source — fork it, run locally, contribute — no vendor lock-in on the editor.
LSPLocal
SSHRemote
AIAgents
Hardware requirements
The app is small. The models are not.
Neural Junkie ships as a ~15 MB desktop app plus an Ollama runtime (~1–2 GB — bundled on macOS, wizard-installed on Linux/Windows). Your first local pull for software development defaults to qwen2.5-coder:14b (~9 GB) and qwen2.5:7b (~4.5 GB) — roughly 15–20 GB total on disk before you chat. The setup wizard reads your RAM and recommends model tags; Settings shows estimated disk and suggested RAM per model.
| Tier | RAM | Developer primary |
|---|---|---|
| minimal | < 8 GB | qwen2.5-coder:7b / llama3.2:3b |
| light | 8–15 GB | qwen2.5-coder:7b |
| recommended | 16–31 GB | qwen2.5-coder:14b |
| heavy | 32 GB+ | 14B + LoRA bases / larger models |
- Multi-agent ≠ multi-model — specialists share one Ollama backend; you pay latency, not six 14B loads.
- Hybrid by design — local iteration, cloud for hard reviews when 14B is not enough.
- Honest limits — collab variance on local models, bounded repo indexes, single local hub — see known issues.
Smarter chat — context that fits the moment
Long conversations need more than a bigger prompt window. Neural Junkie routes intent, scopes workspace attachment, and lets you save facts you want agents to remember.
Routing trace & ReAct tools
Every turn shows routing badges and a live telemetry drawer — model tier, retrieval mode, composer mode. ReAct MCP runs tools on models like Gemma 3 12B without native tool calling.
Chat / Code composer
Auto, Chat, or Code mode on each send. Chat skips heavy tooling for casual turns; Code enables workspace and file workflows. Thread-scoped history keeps side topics focused.
Personal learning
User-confirmed memory with agent, global, and collaboration scopes. Ollama embedding retrieval injects the right notes per turn — nothing saved without your approval.
User rules & find
Per-user markdown rules via Settings/API shape every prompt. In-channel find bar searches message history so you can recover decisions without scrolling forever.
Everything else in the workspace
Beyond IDE v4, local AI, hardware planning, Slack, collaboration, and domain packs — the desktop shell, experts, commands, and provider wiring that tie it together.
Native desktop workspace
Tauri + React app with channels, DMs, threads, workspace tabs, a searchable workspace switcher, and a command palette. Toggle files, code, terminal output, pending changes, and My Agents without leaving the thread.
Custom experts + specialists
Spin up a custom domain expert for any topic via /create-expert or a DM. Enable domain packs for deeper tooling; add repo, Confluence, and CLI agents when you need more.
50+ command actions
Everything from agent lifecycle and provider switches to reindexing, file-change flows, meetings, and collaboration — exposed in the palette with guided forms for complex arguments and slash-form compatibility for automation.
Repo-aware file workflow
Agents propose edits; you approve or reject with diff preview. Local and remote SSH workspaces share the same hub-backed tree, git, and file-change flows.
Model App Store
Toolbar Model library (⇧⌘M): curated Ollama and Hugging Face grid, install progress, Use for agents — extends bundled Ollama after first run.
Runbooks v2
Persisted RunbookDefinition library, run history, connector profiles, and pack-owned templates — import markdown or start from the desktop RB button.
MCP, models, and distribution
MCP export/import for agent knowledge, per-agent providers (Ollama with lifecycle help and an in-app model library, Claude, LM Studio, OpenAI-compatible APIs), optional MCP tool servers, and a Go hub sidecar for single-artifact desktop builds.
See the desktop app
The hub ships a browser chat UI; the Tauri desktop is the full workspace (palette, files, editor, threads). Below is the main workspace; more shots and context live in the README.
How you actually work in it
Same habits as team chat — with agent-native superpowers for individuals and groups.
Routing & context
- @mentions by agent name or type so the right specialist answers.
- Threads for side discussions without spamming the channel.
- Agent review — pull a second opinion by @mentioning another agent in a thread reply.
- Assistant flows — reminders, tasks, notes, meeting ingest, and workspace-grounded follow-ups.
Always-on helpers
- Assistant — command help, product guidance, reminders, tasks, notes, meeting-oriented commands, scoped workspace context, and a safety net when questions sit unanswered.
- First-run setup wizard — get providers and agents configured without reading a novel first.
- Tool approvals — gate risky agent tool use where your stack requires it.
Ship with confidence
Opinionated defaults for teams that want speed and guardrails.
- Single-binary sidecar packaging — ship a .dmg, .msi, or .deb without a fragile pile of services. Remote hosts use the
nj-remotesidecar. - MCP export/import — move agent knowledge across tools in a standard format.
- Auto-updates via the Tauri updater — stay current without chasing tarballs.
- Session persistence on the hub — channels and state come back after restarts (with sensible pruning).
- Markdown & Mermaid in the desktop — diagrams and formatted answers where they help.
- Web + terminal + CLI — same hub speaks to the desktop, browser chat,
make chat, and automation via the Go CLI.
Release notes
Tagged desktop builds and major hub milestones. Install artifacts and checksums are attached to each GitHub release.
Latest download
v1.2.0-beta.22
Semantic stamp dual-gate graduation, continuity dialogue scenarios, and conversation-memory retrieval gates — plus setup wizard / updater fixes from beta.21, one-click Ollama, Neural Canvas + Maps, and everything since beta.6. Release notes.
Download v1.2.0-beta.22 Homebrew install Beta.20 article Release notes CHANGELOG.md
Download the beta
Install from Downloads — macOS: brew install --cask neural-junkie · Linux: brew install neural-junkie (after brew tap camronwood/tap) — or start here for a five-minute walkthrough. Prefer source? Clone the repo and run make start-all.