Local CLI

OpenAI-Compatible CLI Coding Agent for Local & On-Prem LLM Environments

Use your own LLM (vLLM, Ollama, LM Studio, Azure OpenAI, or any OpenAI-compatible API) as a full coding agent - no cloud dependency, no API key costs.

Quick Demo — Tool Usage

Quick Demo: Watch Local CLI use tools to automate real tasks.

https://github.com/user-attachments/assets/77cc96c9-cb22-4411-8744-3a006b00c580

Plan & Execute Demo: Local CLI autonomously editing code with Plan & Execute.

Office Automation Demos

Create professional Word, Excel, and PowerPoint documents with AI:

Why Local CLI?

| Benefit | Description | |---------|-------------| | Zero Cloud Dependency | Runs entirely on your local/on-prem LLM. Your code never leaves your network. | | No API Cost | Use open-source models (Llama, Qwen, DeepSeek, etc.) for free. | | Any OpenAI-Compatible API | Works with vLLM, Ollama, LM Studio, Azure OpenAI, Google Gemini, and more. | | Autonomous Coding Agent | Reads, searches, edits, and creates code files — not just chat. | | Plan & Execute | Breaks complex tasks into TODO steps and executes them step by step. | | Safe by Default | Supervised mode requires your approval before any file modification. | | Office Sub-Agents | Dedicated create/modify sub-agents for Excel, Word, PowerPoint with structured execution. | | Pipe Mode | Non-interactive CLI mode (-p) for scripting and automation pipelines. | | Desktop GUI (Electron) | Dual-window desktop app with chat + real-time task monitoring. Auto-update supported. | | Vision Model Support | Analyze images and screenshots with Vision Language Models. | | Office Automation | Control Excel, Word, PowerPoint directly via PowerShell/COM (Windows). | | Browser Automation | Chrome/Edge CDP control - navigate, click, screenshot, scrape data. |

Installation

CLI (npm)

# Install globally
npm install -g local-cli-agent

# Run
local-cli

The endpoint setup wizard launches automatically on first run.

Desktop App (Windows)

Download the latest LOCAL-BOT-Setup-{version}.exe from the Releases page.

• NSIS installer — installs to %LOCALAPPDATA%\LOCAL BOT\
• Auto-update — the app automatically checks GitHub Releases for updates and notifies you when a new version is available

Build from Source

git clone https://github.com/A2G-Dev-Space/Local-CLI.git
cd Local-CLI
npm install && npm run build
node dist/cli.js

Key Features

Dual-Window Desktop App

The Electron desktop app provides a Chat Window and a separate Task Popup for real-time task monitoring:

• Chat Window - Full-featured chat UI with markdown rendering, code syntax highlighting, and file diff viewer
• Task Popup - Always-on-top task tracker showing current progress, execution status, and tool activity
• Multi-session - Run multiple independent agent sessions in parallel via worker threads
• "Waiting for user input" indicator when the agent needs your response
• Taskbar flashing when tasks complete or user input is needed (even when minimized)
• Auto-expanding input - Text area grows dynamically up to half the screen height
• Table paste - Paste tables from Excel or web pages — auto-converts to markdown
• Image paste - Paste or attach images for Vision model analysis
• Per-model selection - Choose different models for Planning and Execution
• Last folder restore - Reopens the last working directory on app restart
• VSCode diff toggle - Persistent setting for automatic file diff viewing

Plan & Execute

Automatically decomposes requests into TODO steps and executes them:

You: Add a logging system to the project

TODO List                            1/3
  [x] Create logger.ts
  [ ] Add imports to existing code
  [ ] Apply error handling

Deep Research (Search Sub-Agent)

Performs Perplexity-level web research using its own headless Chrome engine — no external search API (Tavily, SerpAPI, Google API) required.

• Dual-engine search: Naver (primary) + Google (secondary)
• Internal source search: Configure additional URLs (Confluence, internal wikis) via researchUrls in config
• Visits actual source pages, extracts content, cross-verifies facts
• Injects today's date for recency assessment
• Returns comprehensive answers with source citations
• Handles Cloudflare-blocked sites gracefully (auto-skip + alternative sources)

Confluence Integration

Edit or create Confluence pages directly via confluence_request:

• Opens a visible browser to access Confluence
• Supports macros, tables, rich text, ProseMirror/TinyMCE editors
• Configure browserServices with type confluence in config to enable

{
  "browserServices": [{ "type": "confluence", "name": "My Confluence", "url": "https://confluence.example.com" }],
  "researchUrls": [{ "name": "My Confluence", "url": "https://confluence.example.com" }]
}

Jira Integration

Manage Jira issues directly via jira_request:

• Opens a visible browser to access Jira (no API key needed)
• Fetch assigned/watching issues via JQL, create issues (Epic, Story, Task, Bug, Sub-task), add comments, transition status
• Two-phase issue creation: fill form → user confirmation → submit
• Autonomous DOM discovery — works with Cloud, Server, and Data Center
• Configure browserServices with type jira in config to enable

{
  "browserServices": [{ "type": "jira", "name": "My Jira", "url": "https://jira.example.com" }]
}

Office Sub-Agents

Office automation uses a dedicated Sub-Agent architecture where each Office app has specialized create and modify agents:

| App | Create Agent | Modify Agent | Capabilities | |-----|-------------|-------------|-------------| | Excel | Structured sheet builder | Tool-based editor | Charts, formatting, conditional formatting, pivot tables, formulas, sparklines | | Word | Section-by-section builder | Tool-based editor | Headers, paragraphs, tables, images, TOC, footnotes, styles, page setup | | PowerPoint | Layout-aware slide builder | Tool-based editor | Slides, text, images, shapes, themes, speaker notes, transitions |

How it works:

1. Enhancement LLM analyzes the user's request and generates detailed content
2. Planning LLM creates a structured JSON plan (design system, sections/sheets/slides)
3. Execution builds the document section-by-section using specialized builder functions
4. Review LLM evaluates quality and triggers refinements if needed

This architecture produces enterprise-quality Office documents with consistent formatting, proper structure, and content appropriate to the topic.

Vision Language Model (VLM)

Analyze images and screenshots directly from the chat:

• Enable via /settings > Vision Model toggle
• Use read_image tool to analyze screenshots, diagrams, and UI mockups
• Supports any OpenAI-compatible Vision endpoint
• Automatic screenshot verification for execution results

Supervised Mode

Every file modification requires your explicit approval:

• Tab to toggle between Auto / Supervised mode
• Only file modification tools need approval (read and search tools are always allowed)
• Reject with feedback to guide the agent's next attempt

Browser Automation

• Navigate pages, click elements, fill forms
• Take screenshots, extract text
• Uses Chrome DevTools Protocol directly — no WebDriver or external server needed

LLM Compatibility

Works well even with smaller or less capable open-source models:

• Schema pre-validation - Fixes malformed tool calls before execution
• Smart retry - Automatically retries on transient LLM errors with context
• Extended retry with user prompt - On persistent failures, offers a retry option instead of crashing
• Loop detection - Breaks out of repetitive tool call cycles
• Tool name sanitization - Cleans up malformed tool names caused by special tokens in model output
• Prompt repetition - Reinforces critical instructions for models with limited context retention

Session Management

• Save and restore conversation history
• Auto-context compression when token usage reaches 80%, preserving TODO state
• Resume work exactly where you left off

Commands & Shortcuts

Slash Commands

| Command | Description | |---------|-------------| | /help | Show help information | | /clear | Reset conversation | | /compact | Compress conversation | | /load | Load saved session | | /model | Switch model | | /settings | Settings menu | | /usage | Token usage | | /docs | Docs management | | /tool | Toggle optional tools (browser/office) |

Keyboard Shortcuts

| Key | Action | |-----|--------| | Ctrl+C | Exit | | ESC | Interrupt current task | | Tab | Toggle Auto / Supervised | | @ | File browser | | / | Command autocomplete |

Configuration

# Terminal — setup wizard launches on first run
local-cli

# Inside Local CLI — open settings menu
/settings

Any OpenAI-compatible API works: vLLM, Ollama, LM Studio, Azure OpenAI, Google Gemini, or internal LLM servers.

Requirements

• Node.js v20+
• npm v10+
• Git
• Windows (required only for Office and Browser automation features)

Changelog

v5.0.4

• Rebranded CLI to local-cli, Electron to LOCAL BOT
• Pipe mode logging setup for --verbose, --debug, --llm-log flags
• SubAgent tool call and phase loggers for -ps (specific) mode

v5.0.2

• Office Sub-Agent v5 — Dedicated create agents with structured execution for Excel, Word, PowerPoint
  • Excel: Sheet-by-sheet builder with automatic chart/formatting generation
  • Word: Section-by-section builder with TOC, page setup, and consistent styling
  • PowerPoint: Layout-aware slide builder with design system and review loop
• Unified branding cleanup (removed all enterprise references)
• Switched to npm publish + GitHub Release deployment (removed binary distribution)
• Electron rebranded to LOCAL BOT with NSIS installer and auto-update via GitHub Release
• Fixed Electron vision tool missing + worker shutdown race condition
• Fixed Jarvis mode auto-update not triggering

v5.0.1

• Added pipe mode (-p) for non-interactive CLI usage in scripts and automation
• Fixed pipe mode final_response capture bug
• Office Sub-Agent architecture (Agent as a Tool) with specialized prompts

v5.0.0

• Sub-Agent architecture — Dedicated sub-agents for Office and Browser with manager/worker LLM pattern
• Jarvis autonomous assistant mode — Always-on-top voice-style UI with tray menu
• Multi-session support — Run multiple independent sessions in parallel via worker threads
• Extended LLM retry — User-facing retry prompt on persistent LLM failures instead of crashing
• Auto-generated session titles — Planning LLM generates meaningful session names
• Planning auto-save to prevent session loss on interruption

v4.5.1

• Fixed Electron shutdown crash (write-after-end)
• Fixed Planning LLM tool_choice fallback for unsupported models
• Fixed Electron chatCompletion response validation crash

v4.5.0

• Electron: 6 UX improvements (waiting indicator, taskbar flash, auto-expanding input, VSCode diff persistence, table paste, image paste)
• Enhanced Planning/Execution prompts + GPT-OSS reasoning_effort support
• Execution result verification rules with Vision screenshot verification
• Auto-compact parity between CLI and Electron (TODO preservation)
• Screenshots now saved to working directory

v4.4.0

• Per-model selection for Planning and Execution LLMs
• Last opened folder restoration on app restart
• Error telemetry system (ErrorReporter)
• Fixed Excel write_range bug (#N/A values)
• Excel tool parameter validation + trim handling

v4.3.0

• Vision Language Model (VLM) support with read_image tool
• Settings UI Vision toggle
• Schema pre-validation + Prompt repetition for less capable LLMs
• Office COM DisplayAlerts auto-suppression
• Office COM Visible always-on + Electron Launch method

v4.2.0

• Electron Dual-window UI (Chat + Task popup) — major UI refactoring
• LLM message structure improvement (XML-based History/Request separation)
• Smart retry + loop detection for LLM robustness
• Tool name sanitization (special token contamination defense)
• edit_file CRLF/LF normalization for Windows files
• PowerShell curl/wget alias auto-substitution
• Supervised Mode bug fixes + escape character handling

v4.1.5 — v4.1.7

• UI/UX improvements and docs-search disable
• Fixed Planning LLM infinite loop (askUser callback)
• 3-second delay + Supervised Mode fixes + model selection bug fix

Documentation

• Developer Guide
• Logging System
• Testing Guide
• Roadmap

Contact

• GitHub Issues: https://github.com/A2G-Dev-Space/Local-CLI/issues
• Email: [email protected]

License

MIT License

GitHub: https://github.com/A2G-Dev-Space/Local-CLI