Forge: A Local-First Autonomous Coding Assistant CLI, the AI Software Engineer in Your Terminal

Forge is a local-first autonomous coding assistant CLI tool positioned as an AI software engineer in your terminal. It adopts an innovative dual-model architecture: the cloud-based "Master Brain" (e.g., Gemini) handles advanced reasoning, project planning, and decision-making, while the local "Local Worker" (e.g., Qwen3.5-9B running via LM Studio) takes care of code generation and rapid iteration. This tool balances capability, speed, and cost, offering multiple modes such as autonomous development, interactive dialogue, and task execution. Its core advantages include data privacy protection, offline capability, low-latency experience, and cost optimization.

Traditional AI coding tools often rely on cloud models, which have issues like privacy risks, network latency, or high costs. Forge's local-first design philosophy aims to address these pain points:

• Data Privacy: Sensitive code stays local; only necessary context is sent to the cloud;
• Offline Capability: Most operations can be performed offline after initialization;
• Cost Optimization: Local models have no API fees, and cloud calls are only used for advanced reasoning;
• Low Latency: Local models respond quickly, avoiding network latency issues.

Dual-Model Architecture

• Master Brain (Cloud): Handles project planning, architecture design, task decomposition, decision-making, and quality control, leveraging the strong reasoning capabilities of cloud-based large models;
• Local Worker (Local): Runs models like Qwen3.5-9B via LM Studio to complete code generation and iterative optimization, ensuring fast response and privacy.

Tech Stack

• Vector Search: Uses FAISS to index project files for quick retrieval of relevant code snippets;
• Dependency Graph: Uses NetworkX to build module relationship graphs, supporting architecture analysis and refactoring;
• Multi-Mode Workflow: Autonomous mode (end-to-end building), interactive mode (REPL dialogue), task mode (specific goal execution).

Installation and Configuration

• Linux/macOS: curl -fsSL https://raw.githubusercontent.com/Tushaarxr/forge/main/install.sh | bash
• Windows: irm https://raw.githubusercontent.com/Tushaarxr/forge/main/install.ps1 | iex
• Environment Requirements: Python3.10+, LM Studio, Gemini API key; complete initialization via forge setup.

Typical Scenarios

• Rapid Prototyping: forge auto "Create a React todo app (TypeScript+Tailwind)";
• Legacy Code Understanding: After forge init, use forge chat to query architecture and code logic;
• Automated Refactoring: forge auto "Convert callbacks to async/await";
• Code Review: Use forge chat to request modification reviews and check security issues.

When using Forge, note the following:

• Model Limitations: Local models (e.g., Qwen3.5-9B) may be insufficient for handling complex logic or rare languages;
• Context Window: For large projects, code must be organized properly to avoid exceeding model capabilities;
• Security Review: Auto-generated code requires manual review, especially for security-sensitive parts;
• Version Control: It is recommended to use Git to enhance change traceability.

Future Directions

• Enhanced Intelligence: Support more cloud/local models (e.g., Llama3 70B) to improve code understanding capabilities;
• Ecosystem Expansion: Develop IDE plugins (VS Code/JetBrains), CI/CD integration, and team collaboration features;
• Specialized Optimization: Optimize for frameworks like React/Django/FastAPI and support domain-specific languages.

Community Contributions

• Local Development: git clone https://github.com/Tushaarxr/forge.git → Install dependencies → Test;
• Contribution Directions: Support more local models, expand language coverage, optimize vector search and dependency graph construction.