llm-box

Turn Natural Language Into Executable Workflows

Agentic Workflow Engine for the Terminal — deterministic execution meets AI agents. Build self-driving workflows with autonomous agent nodes, tool use, and multi-step reasoning.

📋 Table of Contents

🚀 Quick Start
✨ Core Features
🤖 Agent Nodes
📱 HarmonyOS & Mobile Nodes
🔒 Security
🌐 Ecosystem
📚 Documentation
🛠️ CLI Commands
🏗️ Architecture
🗺️ Roadmap
🤝 Contributing
📄 License

🚀 Quick Start

One command to install:

macOS	Linux	Windows
`brew install alib8b8/tap/llm-box`	`curl -fsSL https://raw.githubusercontent.com/alib8b8/llm-box/main/install.sh \| bash`	`irm https://raw.githubusercontent.com/alib8b8/llm-box/main/install.ps1 \| iex`

🌏 China users — use mirror for faster download:

macOS/Linux	Windows
`curl -fsSL https://ghproxy.com/https://raw.githubusercontent.com/alib8b8/llm-box/main/install.sh \| bash`	`irm https://ghproxy.com/https://raw.githubusercontent.com/alib8b8/llm-box/main/install.ps1 \| iex`

Or download from GitCode Releases / GitHub Releases

📖 Interactive Download Page →

Create & run your first workflow

# Generate a workflow from natural language
llm-box create "Summarize today's AI news"

# Or use a built-in template
llm-box create --template research-assistant

# Run it
llm-box run ai-news-summary.yaml

📖 Full Getting Started Guide →

✨ Core Features

Category	Features
Workflow Generation	Natural language → YAML via keyword matching, 100+ built-in templates across 20+ categories
Agent Nodes	30+ AI agent nodes with ReAct, Chain-of-Thought, tool use, and autonomous reasoning
Edge AI Engine	ReAct reasoning loop, 3-tier persistent memory (short/working/long-term), local/cloud model routing, privacy analyzer
Skill Self-Evolution	Agent skills improve with use — auto-tracks success rate, latency, best practices, known pitfalls; auto-optimizes prompts
HarmonyOS Adaptation	Ability launch, atomic service, desktop widget, 7-device-type adaptation (phone/foldable/tablet/TV/car/wearable)
Phone Built-in AI	System event listening (notification/call/SMS/location/battery), on-device LLM inference (1B-8B, INT4/INT8, including SenseNova U1), adaptive power management (eco/balanced/high, battery/thermal-aware), screen understanding, voice input (VAD+wake+ASR), voice output (TTS+voice cloning)
WAIC-Aligned	Blockchain audit trails for agent interoperability, embodied AI robot control (humanoid/arm/drone), L3 smart agent phone capabilities
Cross-Platform Protocol	`intent://` and `ohos://` URI schemes, W3C DID identity verification, cross-domain agent messaging
Ascend NPU Adaptation	7-agent pipeline (search→verify→adapt→quantize→optimize→deploy→doc), CANN/MindIE/MindStudio integration, INT8/FP8 quantization, 1-hour auto-adapt
Code Intelligence	Code graph (AST/call graph/dependency for 158 languages), code knowledge graph (semantic vector retrieval, entity/relation/concept extraction, token-optimized), Codex/OpenCode-compatible tools (glob/grep/list_dir/apply_patch)
Subagent Architecture	Main/sub agent prompt hierarchy (17 specialist templates), borrowed from Grok Build's prompt.md + subagent_prompt.md pattern
Distributed Resilience	Per-node circuit breaker (Closed/Open/HalfOpen state machine), auto-isolation of failing workers, breaker stats endpoint
Privacy by Design	Auto secret redaction (.env/keys/tokens) on file read, outbound data volume monitor with anomaly alerting (prevents Grok-Build-style 27800× leaks)
File Watching	Polling-based file watch node (create/modify/delete events) for log-monitor and file-organizer workflows
TUI Rendering	Terminal Markdown renderer (headings/code/bold/italic/lists/quotes/tables) + Mermaid-to-ASCII converter (flow/sequence diagrams)
Meta Orchestration	Multi-model router (18 models: OpenAI/Anthropic/Google/AndesGPT/SenseNova/DeepSeek/Qwen), 5 strategies (auto/fastest/cheapest/best_quality/privacy_first), hierarchical agent network (supervisor→specialist→worker)
MCP Protocol	MCP bridge client (7 operations, 5 built-in tools) + MCP server mode (HTTP/WebSocket, tool exposure, session management, auth)
Quality Guard	Anti-AI-slop detection, 5 assessment types (AI detection/design/code/writing/overall), auto-fix with quality threshold enforcement
Engineer Skills	16 pre-built skills across 4 domains (frontend/backend/devops/architecture): React/TypeScript/API/Database/CI-CD/Docker/Design Patterns
Skill Distillation	Extract methodologies from books/videos/podcasts/articles into callable skills: workflow/decision/analysis/creative/prompt/checklist
Video Editing	AI video editing: smart_cut/merge/effects/subtitle/storyboard/upscale, 4 styles, 720p/1080p/4k
Utility Nodes	40+ built-in nodes: LLM providers, fetch, execute, transform, file I/O, JSON, notify, condition, combine, call, template
Data & Knowledge	RAG retrieval, knowledge graph extraction/query/traversal, smart model router, multimodal image analysis
Code & Tools	Python code interpreter sandbox, node marketplace, MCP integration, plugin system
Distributed Execution	Coordinator/Worker architecture, horizontal scaling, heartbeat monitoring, circuit breaker
Scheduling	Cron-based scheduled workflows, interval triggers, CLI management
Security	SSRF protection, path traversal prevention, command injection defense, AES-GCM secrets, audit logging, secret redaction, ANSI injection defense, data race fixes, DoS prevention (subscription limits, log rotation, memory caps), 92+ vuln audited
Ecosystem	GitCode G-Star, HarmonyOS Agent Skills, ohpm SDK, OPPO X-OmniClaw Skill Adapter, OPPO 小布技能, AndesGPT, SenseNova
Developer Experience	Web UI editor, workflow visualizer (Mermaid/JSON/DOT/ASCII), TUI with Markdown/Mermaid rendering, 9 languages

🤖 Agent Nodes

Specialized AI agent nodes for autonomous reasoning:

Node	Description
`agent`	General-purpose ReAct agent with tool use, Chain-of-Thought mode
`planner`	Breaks tasks into step-by-step plans
`researcher`	Web research and information gathering
`critic`	Reviews and provides constructive feedback
`evaluator`	Evaluates outputs against criteria
`reflector`	Reflects on process and suggests improvements
`supervisor`	Oversees multi-agent workflows with sequential/parallel/hierarchical/MoE/MindSearch strategies and 16 domain specialists
`code_review`	Automated code review and suggestions
`router`	Routes inputs to appropriate handlers
`human_in_loop`	Pauses for human approval
`meta_orchestrator`	Multi-model router with 18 models, 5 strategies, hierarchical agent network
`code_knowledge_graph`	Semantic code knowledge graph: 158 languages, vector retrieval, entity/relation/concept extraction
`moe_streaming`	MoE expert streaming: consumer hardware runs 744B models with on-demand loading
`cli_session`	Interactive terminal session with context persistence, streaming output, auto-completion
`plugin_system`	Plugin extension: install/uninstall/update from local/git/url/market with sandbox isolation
`mcp_server`	MCP server mode: expose tools via HTTP/WebSocket with session management and auth
`mcp_bridge`	MCP client bridge: 7 operations, 5 built-in tools, protocol-compatible
`quality_guard`	Anti-AI-slop detection: 5 assessment types, auto-fix, quality threshold enforcement
`engineer_skills`	16 pre-built skills: React/TypeScript/API/Database/CI-CD/Docker/Design Patterns
`skill_distill`	Distill methodologies from books/videos/podcasts into callable skills
`voice_output`	TTS + voice cloning: 6 engines, 4 operations, 6 voice styles
`doc_gen`	AI document generation: 7 types (readme/api/function/module/changelog/tutorial/architecture)
`video_edit`	AI video editing: smart_cut/merge/effects/subtitle/storyboard/upscale

HarmonyOS & Mobile Nodes

Node	Description
`harmony_ability`	Launch HarmonyOS Ability (page/slice/service/data) with type validation
`harmony_atomic_service`	Launch atomic service (card-based lightweight app, no install needed)
`harmony_widget`	Manage desktop widgets: add, update, remove, query
`harmony_device_adapt`	Detect 7 device types (phone/foldable/tablet/TV/car/wearable), generate UI adaptation plan
`app_launch`	Launch mobile apps (Android/iOS/HarmonyOS) with platform auto-detection
`ui_automate`	UI automation: click, scroll, type, swipe, screenshot (whitelist-validated actions)
`cross_app_action`	Cross-app workflows: share content, save for later, compare prices
`intent_router`	Route user intents to appropriate handlers with domain classification
`device_state`	Query device state: battery, network, location, apps, storage
`agent_message`	Send cross-domain messages between agents using W3C DID identity
`agent_inbox`	Query and manage agent message inbox
`system_event`	Listen for mobile system events: notification, call, SMS, location, battery, alarm, screen state
`ondevice_llm`	Run LLM inference locally on device (1B-7B models, INT4/INT8/FP16, llama.cpp/MLC-LLM/ONNX backends)
`power_manager`	Adaptive power management: eco/balanced/high profiles, battery-aware & thermal-aware throttling
`blockchain_audit`	Record workflow execution on blockchain for tamper-proof audit trails (Ethereum/Hyperledger/simulated). Aligns with WAIC Agent Interoperability Initiative
`screen_understanding`	L3-level screen content understanding: parse UI elements, identify actionable items, generate interaction plans for agent phones
`voice_input`	Voice pipeline: VAD, wake word detection (hey_box/hello_box/hi_box/ok_box/box_box), speech-to-text with on-device support
`robot_control`	Plan and execute robot action sequences for embodied AI: humanoid/mobile_base/arm/drone/dog/wheelchair with safety checks
`andesgpt`	OPPO AndesGPT integration: Tiny (端侧1B) / Turbo (端云协同7B) / Titan (云端100B+), PersonaX personalization, end-cloud collaboration

Ascend NPU Adaptation Nodes

Node	Description
`ascend_model_search`	Search models in AtomGit/HuggingFace ModelZoo
`ascend_model_verify`	Verify model manifest, dependencies, License, Ascend compatibility
`ascend_model_adapt`	Adapt model to Ascend NPU via msTransplant, handle operator patches
`ascend_model_quantize`	INT8/FP8/W8A8 quantization via msModelSlim with accuracy comparison
`ascend_model_optimize`	Performance tuning via msProf/msprof-analyze, bottleneck analysis
`ascend_model_deploy`	MindIE Service deployment, OpenAI API compatibility test
`ascend_model_doc`	Auto-generate benchmark report and reproduction guide
`ascend_model_agent`	End-to-end orchestrator (mode: full/quick/tune)

Code Intelligence & Tool Compatibility Nodes

Node	Description
`code_graph`	Extract code structure (imports/functions/calls) for Go/Python/JS/TS, output JSON or Mermaid graph
`file_watch`	Watch a path for file changes (create/modify/delete), polling-based, context-aware
`glob`	Recursive file glob matching (`*/.go`), depth-limited, Codex-compatible
`grep`	Recursive content search with regex, binary-skip, Codex/OpenCode-compatible
`list_dir`	List directory contents (optional recursive), Codex-compatible
`apply_patch`	Apply unified diff patches atomically (validate-then-commit), Codex-compatible

🔒 Security

llm-box takes security seriously. Key protections:

Protection	Implementation
SSRF Protection	Custom `DialContext` validates IPs at connect time (prevents DNS rebinding)
Path Traversal	Input validation + symlink resolution, all paths confined to working directory
Command Injection	Allowlist mode blocks shell metacharacters; safe mode disables execution entirely
Secrets Management	AES-GCM encryption with PBKDF2 (600K iterations), file permissions `0600`
Timing Attack	`subtle.ConstantTimeCompare` for authentication tokens
Fail-Closed Auth	Empty token = request rejected (503)
Audit Logging	All commands logged with redacted secrets, `0600` permissions
DID Identity	W3C DID format validation, signature verification, cross-domain message auth
Memory Safety	Layered memory with LRU eviction, symlink-resistant persistence, 0600 file perms
Prompt Injection	Skill evolution sanitizes best practices/pitfalls before injecting into prompts
Secret Redaction	Auto-detect & mask 10+ secret patterns (AWS/GitHub/Slack/JWT/private keys) on file read; `.env`/credentials fully masked by default
Outbound Monitoring	Sliding-window data volume monitor with anomaly alerting (prevents Grok-Build-style 27800× data leaks)
Circuit Breaker	Per-worker breaker (Closed→Open→HalfOpen), auto-isolation of failing nodes prevents cascade failures
Atomic Patches	`apply_patch` validates-then-commits with temp staging + atomic rename; no partial writes on failure
ANSI Injection	TUI Markdown/Mermaid renderers strip terminal control sequences (CSI/OSC/DCS) from user input
Tool Portability	Codex/OpenCode-compatible tools (glob/grep/list_dir/apply_patch) with full path/symlink/DoS hardening
Session Limits	CLI sessions auto-expire after 24h (max 500), MCP sessions capped at 1000 with cleanup
Plugin Limits	Max 100 plugins, HTTPS-only URLs, restricted git hosts (GitHub/GitLab/GitCode/Gitee)
Resource Limits	Code knowledge graph: max 5000 files/depth 5; video edit: shell metacharacter filtering
Concurrent Safety	Per-session rand mutex, RWMutex for shared state, no global mutable state without locks

📖 Security Guide → | Audit Logs →

🌐 Ecosystem

llm-box participates in multiple open-source ecosystems:

Ecosystem	Status	Description
GitCode G-Star	Applied	Compute support, traffic exposure, HarmonyOS certification
HarmonyOS Agent Skills	Published	8 skills: ability launch, atomic service, widget, device adapt, cross-app, agent message, intent router, device state
ohpm SDK	Published	`@llm-box/workflow-engine` — ArkTS SDK with WorkflowEngine, 30+ node types, device adaptation, intent protocol
Ascend NPU Adaptation	Active	7-agent auto-adapt pipeline, 3 workflow templates (end-to-end/quick/performance-tune), CANN/MindIE integration
SenseNova	Active	API integration (6 models), on-device U1-Lite support (8B/A3B MoE), 8 skills for SenseNova ecosystem
GitHub	Active	CI/CD, CodeQL security scan, automated releases

HarmonyOS Device Support

Device Type	Key Capabilities
Phone (Standard)	touch, camera, gps, nfc, biometrics
Phone (Dual Fold)	foldable_screen, multi_window, drag_to_split
Phone (Triple Fold)	foldable_screen, multi_window, drag_to_split
Tablet	stylus, multi_window, split_screen
Smart Screen	voice, gesture, remote_control
Car	steering_wheel_control, hud, voice
Wearable	heart_rate, accelerometer, gyroscope

Ascend NPU Hardware Support

Hardware	Position	Model Scale
Ascend 910B	Training/Inference	7B-70B
Ascend 910C	Training/Inference	7B-170B
Atlas 800I A2	Inference Server	7B-70B
Atlas 300I Duo	Edge Inference	<13B
310P	Edge Inference	<7B

📖 G-Star Application → | HarmonyOS Skills → | ohpm SDK → | Ascend Adaptation →

📚 Documentation

Getting Started

Getting Started Guide — Installation, first workflow, next steps
Examples — 10 ready-to-use workflow templates

Core Concepts

Dataflow & Variables — Step-to-step data passing, {{input}}, {{step.N}}, {{var.NAME}}, {{secret.GROUP.KEY}}
Distributed Execution — Coordinator/Worker setup, configuration, scaling
Scheduling — Cron workflows, schedule management
MCP Integration — Connect external tools via Model Context Protocol
Plugins — Install and manage community plugins

Advanced

Web UI Editor — Visual workflow builder
Visualizer — Mermaid/JSON/DOT/ASCII diagrams
Tenant Isolation — Multi-tenant workspace separation
Custom Nodes — Build nodes in any language
Troubleshooting — Error codes, common issues, FAQ

Reference

🛠️ CLI Commands

llm-box create [description]    Generate workflow from description
llm-box run <file>             Run a workflow
llm-box secrets add            Store an encrypted secret
llm-box secrets list           List secrets in a group
llm-box schedule create        Create a scheduled workflow
llm-box schedule list          List scheduled workflows
llm-box coordinator            Start distributed coordinator
llm-box worker                 Start distributed worker
llm-box ui                     Start web UI editor
llm-box visualize <file>       Visualize a workflow
llm-box validate <file>        Validate a workflow file
llm-box node install           Install external node
llm-box plugin install         Install a plugin
llm-box version                Show version
llm-box help                   Show full help

📖 Full CLI Reference →

🏗️ Architecture

┌─────────┐     ┌─────────┐     ┌──────────────┐     ┌──────────┐     ┌────────┐
│  Prompt │────▶│ Planner │────▶│ Workflow YAML│────▶│ Executor │────▶│ Result │
└─────────┘     └─────────┘     └──────────────┘     └──────────┘     └────────┘
                                                          │
                                              ┌───────────┴───────────┐
                                              ▼                       ▼
                                      ┌──────────────┐         ┌──────────────┐
                                      │  Agent Nodes │         │ Utility Nodes│
                                      │  ReAct Loop  │         │ fetch, exec  │
                                      └──────────────┘         └──────────────┘

Key components:

Generator — Keyword-based workflow generation from natural language, 100+ ready-to-use templates
Parser — YAML workflow validation and parsing
Executor — Deterministic step execution with dependency tracking
Expression Engine — Variable substitution, secrets injection, file reading
Registry — 50+ built-in nodes plus external node discovery and loading
Edge Router — ReAct reasoning loop, 3-tier persistent memory, local/cloud model routing
Skill Evolution — Self-improving agent skills with success rate tracking and prompt optimization
Intent Protocol — intent:// and ohos:// URI schemes, W3C DID identity, cross-domain messaging
Ascend Adaptation — 7-agent pipeline for Ascend NPU model adaptation (search/verify/adapt/quantize/optimize/deploy/doc)
Code Intelligence — Code graph extraction, Codex/OpenCode-compatible tool nodes (glob/grep/list_dir/apply_patch)
Subagent Prompts — 17 specialist prompt templates, main/sub agent hierarchy (Grok Build pattern)
Circuit Breaker — Per-worker Closed/Open/HalfOpen state machine for distributed resilience
Privacy Layer — Secret redaction on file read, outbound data volume anomaly monitor
Coordinator/Worker — Distributed task scheduling and execution with circuit breaker protection

🗺️ Roadmap

Version	Status	Features
v0.1	✅ Released	Core workflow engine, 10 utility nodes
v0.2	✅ Released	LLM nodes, MCP integration, external nodes
v0.3	✅ Released	Agent nodes, distributed execution, Web UI, scheduling
v0.4	✅ Released	Code interpreter, RAG, knowledge graph, smart router, multimodal, node marketplace, 100+ templates, 16 specialists, Chain-of-Thought
v0.5	✅ Released	ReAct engine, layered memory, skill self-evolution, HarmonyOS adaptation (7 device types), cross-platform protocol (intent:// + ohos://), W3C DID identity, cross-domain agent messaging, GitCode G-Star + ohpm ecosystem
v0.5.1	✅ Released	Ascend NPU adaptation (7-agent pipeline, 3 workflow templates, CANN/MindIE integration)
v0.5.2	✅ Released	Grok Build-inspired capabilities: code graph, subagent prompt hierarchy, circuit breaker, secret redaction, file watch, Codex/OpenCode tool compat, TUI Markdown/Mermaid rendering (15-vuln audited)
v1.0	📅 Q3 2026	Stable API, full documentation, LTS

📖 Full Roadmap →

🤝 Contributing

We welcome contributions from the community!

Good First Issues

New to the project? Start here:

🐛 Bug fixes labeled "good first issue"
📝 Documentation improvements
✅ Add test coverage for low-coverage packages
🔧 New utility nodes (see custom nodes guide)
🌐 Add or improve translations (i18n)

How to Contribute

Fork the repository
Create a branch: git checkout -b feature/your-feature
Make changes and add tests
Run tests: go test ./...
Submit a pull request

Review Process

CI must pass (build, tests, lint, security scan)
At least one approval from a CODEOWNER
Code follows Go conventions (gofmt, go vet)
New features include tests and documentation

📖 Full Contributing Guide →

📄 License

MIT License — see LICENSE for details.

Built with ❤️ for developers who love the terminal

GitHub · Issues · Discussions

Name		Name	Last commit message	Last commit date
Latest commit History 321 Commits
.claude-plugin		.claude-plugin
.codex-plugin		.codex-plugin
.github		.github
.grok-plugin		.grok-plugin
.grok/skills		.grok/skills
.trae/skills/llm-box-workflow		.trae/skills/llm-box-workflow
.traecli/skills/llm-box-workflow		.traecli/skills/llm-box-workflow
cmd/llm-box		cmd/llm-box
commands		commands
contrib/openclaw		contrib/openclaw
docs		docs
ecosystem		ecosystem
examples		examples
grok-mcp-server		grok-mcp-server
internal		internal
nodes		nodes
skills		skills
templates		templates
vscode-extension		vscode-extension
.codexignore		.codexignore
.gitignore		.gitignore
.goreleaser.yml		.goreleaser.yml
.gosec.json		.gosec.json
.mcp.json		.mcp.json
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
GEMINI.md		GEMINI.md
LICENSE		LICENSE
LOOP.md		LOOP.md
Makefile		Makefile
README.md		README.md
SECURITY.md		SECURITY.md
STATE.md		STATE.md
gemini-extension.json		gemini-extension.json
generate_templates.go		generate_templates.go
go.mod		go.mod
go.sum		go.sum
install-skill.sh		install-skill.sh
install.ps1		install.ps1
install.sh		install.sh
logs.zip		logs.zip
loop-budget.md		loop-budget.md
loop-constraints.md		loop-constraints.md
loop-run-log.md		loop-run-log.md
nodes-registry.json		nodes-registry.json

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

llm-box

📋 Table of Contents

🚀 Quick Start

Create & run your first workflow

✨ Core Features

🤖 Agent Nodes

HarmonyOS & Mobile Nodes

Ascend NPU Adaptation Nodes

Code Intelligence & Tool Compatibility Nodes

🔒 Security

🌐 Ecosystem

HarmonyOS Device Support

Ascend NPU Hardware Support

📚 Documentation

Getting Started

Core Concepts

Advanced

Reference

🛠️ CLI Commands

🏗️ Architecture

🗺️ Roadmap

🤝 Contributing

Good First Issues

How to Contribute

Review Process

📄 License

About

Uh oh!

Releases 6

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

llm-box

📋 Table of Contents

🚀 Quick Start

Create & run your first workflow

✨ Core Features

🤖 Agent Nodes

HarmonyOS & Mobile Nodes

Ascend NPU Adaptation Nodes

Code Intelligence & Tool Compatibility Nodes

🔒 Security

🌐 Ecosystem

HarmonyOS Device Support

Ascend NPU Hardware Support

📚 Documentation

Getting Started

Core Concepts

Advanced

Reference

🛠️ CLI Commands

🏗️ Architecture

🗺️ Roadmap

🤝 Contributing

Good First Issues

How to Contribute

Review Process

📄 License

About

Topics

Resources

License

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 6

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages