BLAXIN

An open-source AI desktop agent for Linux — it plans multi-step tasks and actually does them on your computer.

v1.4.0 is the final public release. The repository may receive documentation, maintenance, and non-release improvements, but no subsequent public version is planned.

Releases · Install · Getting started · Capabilities · Architecture · Troubleshooting

BLAXIN is a desktop AI agent that takes a plain-language instruction — "organize my downloads folder", "open example.com", "kill process 4321" — and carries it out on a real Linux machine: driving the terminal, the file system, the GUI, and the browser through Google Chrome.

Instead of only suggesting commands, BLAXIN plans the steps, executes them through a set of auditable tools, verifies every outcome against the real environment (file read-backs, page transitions, process tables, screenshots), and reports honest results — including failures. High-impact actions such as deletions and destructive shell commands pause the agent and require your explicit approval first.

Everything runs locally: the agent, its memory, and — if you choose — the AI model itself (via Ollama). Cloud providers are supported but never required.

_{The Jarvis command HUD — a real screenshot of BLAXIN v1.4.0 running on Linux. More below.}

• Desktop control — mouse, keyboard, window management via xdotool; screen-bounds grounding; verified screenshots with OCR perception
• Browser automation — real Chrome over CDP: navigation, tabs, multi-field form fill with per-field read-back, verified form submission, disk-verified downloads
• File operations — read/write with verified read-back, organize-by-type, batch move/copy/rename, and content-hash (SHA-256) duplicate detection and dedupe
• Process control — list, inspect, and kill by explicit pid, with zero signals on refusal and kills verified by fresh read-back
• Missions & specialists — persistent multi-step missions with per-step checkpoints, bounded specialist agents, deterministic recovery and re-planning
• Memory & learning — layered memory (failures / environment / episodes / procedures); verified multi-step runs become reusable procedures, stale ones auto-rollback
• System control — real telemetry (CPU/RAM/disk/network/battery), volume control via PipeWire, clipboard, system info
• Voice — browser speech recognition and synthesis (feature-detected)

Every capability claim is grounded in the code and test suite — see docs/CAPABILITIES.md for the full detail, including per-capability verification.

The one-command installer downloads the latest stable release from GitHub — currently v1.4.0, the final public release — verifies its SHA-256 checksum, and installs it with desktop integration:

curl -fsSL https://raw.githubusercontent.com/tasinxxx/Blaxin/main/blaxin/install.sh | bash

The installer is distro-aware:

• Debian / Ubuntu / Kali / Mint — installs the .deb package (runs against your system WebKitGTK — the most reliable configuration)
• Other distros — installs the portable AppImage to /opt/blaxin with a desktop entry
• Force the AppImage anywhere: ... | bash -s -- --appimage

# .deb (Debian/Ubuntu/Kali)
sudo apt install ./blaxin_1.4.0_amd64.deb
blaxin

# AppImage (portable, other distros)
chmod +x BLAXIN_1.4.0_amd64.AppImage
./BLAXIN_1.4.0_amd64.AppImage

cd blaxin
docker compose up -d

git clone https://github.com/tasinxxx/Blaxin.git
cd Blaxin/blaxin
./start.sh          # installs deps, starts backend :3001 + client :5173

1. Start BLAXIN from your application menu, or run blaxin in a terminal.
2. The setup wizard walks you through choosing an AI provider:
   • Cloud — pick OpenRouter (recommended for model variety), OpenAI, Anthropic, Google, Groq, or Together and paste an API key. Keys are AES-256-CBC encrypted locally and never sent to the UI or logs.
   • Local — install Ollama (BLAXIN can install and start it for you), pull a model, and BLAXIN recommends the best catalog model that actually fits your CPU/RAM/GPU.
3. Type an instruction in the composer and watch the plan, execution, and verification in real time.

No API key yet? Deterministic commands still work (try /status or list /tmp) — the full AI reasoning loop needs a provider or local model.

Natural language goes straight into the composer:

list the contents of /tmp
organize ~/Downloads by type
open https://github.com/tasinxxx/Blaxin/releases
find duplicate files in ~/Pictures
what processes are running
set volume to 40

Slash commands: /help /status /queue /missions /memory /clear /stop /version

Every high-impact action (deletes, destructive commands, browser/desktop actions) opens an approval dialog before executing — you choose approve-once, approve-for-task, approve-for-session, or deny.

User (Jarvis HUD / voice)
        │
Jarvis Command Layer — deterministic router, slash commands, task queue
        │
Agent Orchestrator (WebSocket) — planning, specialist delegation,
        │                        recovery/re-planning, verification
        │
Provider Abstraction — OpenRouter · OpenAI · Anthropic · Google ·
        │               Groq · Together · Ollama (local) · OCI tunnel
        │
Tools — terminal · filesystem · bulk-files · computer-control ·
        │  process-control · browser (CDP) · blaxin_web · screenshot ·
        │  clipboard · search · system-info · system-audio
        │
Verification-in-depth — read-backs, page transitions, process tables,
        │                 screenshots · evidence recorded in the Mission Journal
        ▼
Your Linux desktop

A deeper walkthrough — components, data flow, and the verification/recovery loop — is in docs/ARCHITECTURE.md.

Honest limitations of the current release:

• Linux only, x86_64 only — no Windows or macOS packages; ARM64 is not supported by the installer
• Wayland is partial — desktop control works, but some verification read-backs (e.g. pointer position) are unavailable there and are reported as such rather than guessed
• AppImage blank-window bug — the bundled WebKitGTK can fail EGL initialization on some hosts (upstream Tauri issue); use the .deb on Debian-family distros. See troubleshooting.
• Vision-model routing — screenshot-driven LLM decisions need a vision-capable model; without one, grounding falls back to OCR and the result says so
• No WhatsApp/cloud-automation integrations — deliberately out of scope for a privacy-first desktop agent
• Voice — browser-based STT/TTS depends on your browser and audio hardware

• High-impact actions pause the agent and require explicit user approval (once / per-task / per-session scopes)
• Filesystem writes and deletes are hard-blocked on system-critical and credential paths (/etc, /boot, ~/.ssh, key files, …)
• API keys are AES-256-CBC encrypted at rest with 0600 permissions, never logged, never sent to the frontend
• All WebSocket and state-changing HTTP requests are origin-checked; the desktop app binds the backend to 127.0.0.1 only
• Secrets are refused by the memory store and masked in all logs

Full detail: SECURITY.md. To report a security vulnerability, please follow the responsible disclosure policy — do not open a public issue.

v1.4.0 is the final public release of BLAXIN. The repository may still receive documentation, maintenance, and non-release improvements, but no subsequent public version is planned.

BLAXIN v1.4.0 is complete and remains the recommended version for all users. It ships the Jarvis HUD, mission coordination, verified browser/computer/file/process control, and the layered memory system — all backed by a 900+ test suite and real-machine runtime probes recorded in the release notes.

• Download: v1.4.0 release — the final public release
• Updater manifest: blaxin/update/latest.json (signed artifacts)
• Maintenance: bug reports and security reports are welcome via the issue tracker; documentation and non-release maintenance may continue, but no further numbered release is planned

All screenshots are real captures of the running application — regenerated with blaxin/e2e/scripts/capture-screenshots.mjs.

Agent terminal — live event stream and slash-command composer	Confirmation gate — explicit approval before a high-impact action
Verified mission — a real bulk-files task settled with per-step verification

Contributions are welcome — bug reports, documentation improvements, and code. See CONTRIBUTING.md for the workflow and project conventions.

• Bug report → open an issue (the journal excerpt from the Mission Journal page makes bugs much easier to diagnose)
• Feature idea → feature request
• Security vulnerability → SECURITY.md (private disclosure, please)

BLAXIN is released under the MIT License.