0xAgent — Autonomous AI Developer & Web-IDE Platform

Next-generation autonomous AI coding platform and Web-IDE with a built-in local inference engine (llama.cpp), full zero-config agent harness pipeline, and hybrid cloud fallback.

Quick Start • Terminal CUI • Key Features • Architecture • Configuration • License

English • Русский

---

---

⚡ 1-Click Quick Start

Install and configure 0xAgent in a single command. The installer automatically verifies prerequisites (Node.js LTS, Git), sets up SSL, builds the client, compiles the native Windows tray host, and binds the global 0xagent CLI.

Windows (PowerShell)

irm https://raw.githubusercontent.com/T58574/0xAgent/main/install.ps1 | iex

Linux / macOS / WSL (Bash)

curl -fsSL https://raw.githubusercontent.com/T58574/0xAgent/main/install.sh | bash

---

🎮 Terminal CUI & CLI Agent

0xAgent provides a complete, rich Terminal CUI (Character User Interface) and CLI execution runner (similar to Claude Code, Antigravity CLI, Codex CLI, and OpenCode CLI) that can be invoked from any directory or automated from scripts:

# Launch full interactive CUI REPL session in the current directory:
0xagent

# Interactive Settings TUI (toggle Auto-Open Browser, Permissions, Themes, etc.):
0xagent settings

# Interactive Model TUI (with horizontal reasoning effort slider and GGUF scanning):
0xagent model

# Interactive Persona Switcher:
0xagent persona

# One-shot prompt execution directly in your terminal:
0xagent "Analyze this project and suggest architectural improvements"
0xagent -p "Fix type errors in server/agent.ts"

# Quiet output for script automation & piping (e.g. CI / Headless scripts):
0xagent -p "Summarize git diff" --quiet
git diff | 0xagent "Explain these changes"

# Manage local llama-server:
0xagent server status
0xagent server start
0xagent server stop
0xagent server logs
0xagent server purge

# View and update configuration:
0xagent config show
0xagent config set permission_preset unrestricted
0xagent config set auto_open_browser true

# Service supervision:
0xagent start           # Launch background System Tray host (Windows)
0xagent status          # Check backend health & telemetry
0xagent purge-vram      # Force release GPU VRAM
0xagent update          # Pull latest releases & rebuild
0xagent stop            # Terminate all processes

⌨️ In-CUI Slash Commands & Interactive Modals

When running 0xagent, type / to bring up the command picker or use slash commands directly:

Slash Command	Description
/settings	Open interactive 2-column settings table with search filter (Auto Open Browser, Permission Preset, Reasoning Effort, Theme, etc.)
/model	Open model picker with horizontal Effort Slider (low / medium / high / auto), local GGUF scan, and cloud models
/persona	Open interactive persona profile selector
/server [action]	Manage local llama-server (start, stop, status, logs, purge)
/status	View system hardware, GPU VRAM usage, and active session telemetry
/compact	Trigger 4-tier context compaction and token pruning on demand
/config [key] [val]	Inspect or update configuration key inline
/clear, /reset	Clear conversation history and reset context checkpoint
/help	Display command guide and keybindings
/exit, /quit	Exit the CUI session

Tip

No Browser Popups on Startup: By default, auto_open_browser is set to false. Starting 0xAgent never interrupts your workflow with an unwanted browser tab. You can toggle this anytime in /settings, the Web IDE settings, or via 0xagent config set auto_open_browser true.

---

🚀 Key Features & Agent Harness

Unlike conventional wrappers requiring external servers (like Ollama or vLLM), 0xAgent is the first all-in-one platform featuring a native, built-in inference supervisor, a standalone Telegram bot, a persistent SQLite memory engine, and a production-grade autonomous agent harness out of the box with zero complex setup.

📱 Standalone Telegram Bot Subsystem

• Direct Local Model Querying: Built with GrammY, the bot connects directly to the local llama-server (127.0.0.1:11434), delivering 100% private, cloud-free inference.
• Hardware-Enforced Security Whitelist: Strict Telegram ID filtering (telegram.whitelist) ensures only authorized user IDs can query the bot.
• Multi-Turn Context & Smart Splitting: Retains conversational context with automatic window management and converts Markdown into clean Telegram HTML cards with syntax highlighting, blockquotes, and balanced message chunking (<4096 chars).
• Built-in Commands: /start, /help, /status (server & model health check), /model (active LLM inspection), /reset, /clear (context flush).

🧠 Memory Engine v1.0 & Scoped SQLite Architecture

• Native SQLite WAL Store: All memories, audit trails, and episodes persist in ~/.0xagent/memory.db with native node:sqlite speed and FTS5 full-text indexing.
• Dual-Scope Isolation: Strict physical partition between global user preferences and project-specific knowledge (project_id), resolving workspace path aliases automatically.
• Dynamic Token Budget Allocator: Automatically scales context memory injection (0..400 tokens) — allocating 0 memories during casual dialogue to maximize KV cache throughput.
• Memory Decay & Conflict Resolution: Automatically decays confidence scores over time, archives stale memories (< 0.1), and supersedes duplicates deterministically.
• Dynamic USER.md Compiler: Compiles living user preferences directly into the system prompt with zero manual file editing.

⚡ Dual Inference & Unified Model Hub

• Native llama-server Supervisor: 1-click binary downloader and automatic GPU layer offloading (-ngl), Flash Attention (-fa on), quantized KV cache (-ctk q8_0 -ctv q8_0), and automated VRAM release when idle or switching to cloud.
• Local GGUF Model Hub: Direct zero-config support for Qwen 2.5 Coder, Gemma 4, DeepSeek, and Llama 3.3.
• Hybrid Cloud Fallback: Instant toggle to Google AI Studio (Gemini 3.7/3.6/3.1 Pro, Flash Lite) with inline reasoning effort toggles (off, low, medium, high).
• LAN Remote Workstation Mode: Run the lightweight Web IDE / CUI on an ultrabook (~150 MB RAM) while offloading heavy LLM inference to a dedicated GPU workstation on your local network.

🛠 Production-Grade Zero-Slop Agent Harness

• Concurrent Tool Execution: Read-only exploration tools (read_file, list_dir, grep_search, fff_search, web_search) execute in parallel via Promise.all(), speeding up repository scans by 3-5x.
• Whitespace-Tolerant Patching (patch_file): Robust multi-chunk search/replace block patcher ensuring surgical edits with zero data loss or file truncations.
• Sandboxed Code Mode (<code_run>): In-memory VM runtime enabling the agent to execute complex Node.js automation scripts with async host tool bindings in a single turn.
• Two-Tier Approval Protocol: Non-blocking intent suggestions (<quick_replies>) and cryptographic approval gates (<request_approval>) with SHA-256 validation for destructive operations.
• Self-Improvement & Staged Proposals (selfPatchEngine.ts): Core system modifications are automatically intercepted, isolated into staged proposals, and can be reviewed, applied, or rolled back safely.
• Oscillation & Loop Breaker (loopBreaker.ts): 8-step rolling history tracking with canonical argument sorting, preventing repetitive tool cycling.
• 4-Tier Context Compaction (compactionPipeline.ts): Coordinated token optimization featuring zero-token tool pruning with error retention, CoT thought stripping, bounded windowing, and milestone summarization.
• Output Spiller (outputSpiller.ts): Automatically offloads massive terminal outputs (>24 KB) to disk (~/.0xagent/spill/*.log) to shield the LLM context window.
• Privacy Web Search & Fast File Finder: Local SearXNG / DuckDuckGo web research with Markdown scrapers and sub-3ms Rust-accelerated file finder (@ff-labs/fff-node).

---

🏛 Architecture

System Flow & Component Topology

flowchart TD
    subgraph UI ["Frontend Web IDE (React 19 + TypeScript + Tailwind 4)"]
        Chat["Chat & Reasoning Stream (<think>)"]
        Editor["Monaco Code Editor & Tabs"]
        PlanHUD["Live Plan Progress HUD (todo_write)"]
        CmdBar["Floating Command Bar & Permission Matrix"]
    end

    subgraph Host ["Zero-Overhead Supervisor & Host"]
        Tray["Native C# Tray Launcher (0xAgent.exe)"]
        CLI["Universal CLI Hub (0xagent)"]
    end

    subgraph Core ["0xAgent Backend Engine (Express + WebSocket)"]
        AgentLoop["Agent Orchestrator Loop (agent.ts)"]
        Compactor["4-Tier Context Compactor & Pruner"]
        LoopGuard["Loop Breaker & Output Spiller"]
        Sandbox["Code Mode VM Sandbox (<code_run>)"]
        Dispatcher["Parallel Tool Dispatcher"]
    end

    subgraph Inference ["Dual Inference Engine"]
        LlamaSup["llama-server Supervisor\n(GGUF / Flash-Attn / VRAM Purge)"]
        CloudAPI["Cloud LLM Gateway\n(Gemini 3.6 / Flash Lite / Groq)"]
    end

    subgraph Tooling ["Workspace Tooling & External Services"]
        FilePatcher["Fuzzy Multi-Chunk Patcher"]
        FFF["Rust Fast File Finder (FFF)"]
        Terminal["Live Terminal Supervisor"]
        Search["SearXNG / DuckDuckGo Engine"]
    end

    Tray --> UI
    CLI --> Core
    UI <===>|"HTTPS REST & Duplex WSS"| Core
    Core --> AgentLoop
    AgentLoop --> Compactor
    AgentLoop --> LoopGuard
    AgentLoop --> Sandbox
    AgentLoop --> Dispatcher
    Dispatcher --> FilePatcher
    Dispatcher --> FFF
    Dispatcher --> Terminal
    Dispatcher --> Search
    AgentLoop <===> Inference
    Inference --- LlamaSup
    Inference --- CloudAPI

Loading

High-Level Topology Schema

┌────────────────────────────────────────────────────────────────────────┐
│                      0xAgent Web IDE Interface                         │
│       (React 19 + Vite 7 + Monaco Editor + Glassmorphism Theme)        │
└───────────────────────────────────┬────────────────────────────────────┘
                                    │ HTTPS REST & Duplex WSS
┌───────────────────────────────────▼────────────────────────────────────┐
│                    0xAgent Backend Engine & Harness                    │
│   ┌────────────────────────────────────────────────────────────────┐   │
│   │ Agent Loop • 4-Tier Context Compaction • Loop Breaker • Sandbox │   │
│   └───────────────────────────────┬────────────────────────────────┘   │
└───────────────────┬───────────────┴───────────────┬────────────────────┘
                    │                               │
┌───────────────────▼──────────────┐ ┌──────────────▼────────────────────┐
│  Built-in llama.cpp Supervisor   │ │      Hybrid Cloud Gateway         │
│  (Native GGUF / GPU Offloading)  │ │   (Google AI Studio / Groq API)   │
└──────────────────────────────────┘ └───────────────────────────────────┘

---

🌐 Localization

0xAgent provides full out-of-the-box bilingual support for English and Russian (Русский) across the entire interface, tool outputs, settings, and voice telemetry. Toggle instantly via the [EN] / [RU] badge in the navigation bar or configure via 0xagent config.

• 📖 Документация на русском языке доступна в README.ru.md.

---

📁 Configuration

All runtime configurations, model weights, personas, and memory are stored in ~/.0xagent/:

Directory / File	Description
~/.0xagent/config.json	Global settings, API keys, active models, proxies, and security permissions
~/.0xagent/memory.db	Canonical SQLite database (WAL mode: memories, episodes, relationships, FTS5)
~/.0xagent/certs/	Local SSL development CA and self-signed certificates for HTTPS / WSS
~/.0xagent/models/	Local GGUF model files repository
~/.0xagent/llama/	Managed llama-server.exe binary builds
~/.0xagent/personas/	System personas & memory (SOUL.md, USER.md, TOOLS.md)
~/.0xagent/sessions/	Dialogue session history and branching checkpoints
~/.0xagent/workspaces/	Isolated workspace sandbox directories
~/.0xagent/spill/	Disk spilled logs for outputs exceeding 24 KB

---

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.