Forge

Lightweight coding agent for local and hosted LLM workflows.

Forge is a local-first coding assistant designed for developers who want transparent tool execution, persistent sessions, and full control over their workflow.

Features

• Multiple providers
  • Ollama
  • Gemini
  • OpenRouter
• Agent and non-agent modes
• Interactive chat sessions
• Persistent SQLite-backed history
• Workspace-aware tools
• File search and editing
• Shell command execution
• Git integration
• Code review command
• Model usage tracking
• Tool execution tracking
• Approval mode for destructive actions

Installation

git clone https://github.com/mm4cN/forge.git
cd forge

python -m venv .venv
source .venv/bin/activate

pip install -e .

Ollama

ollama serve
ollama pull qwen2.5-coder:7b

Gemini

export GEMINI_AUTH_KEY="..."

Quick Start

Ask a model:

forge ask "Explain RAII in C++"

Run the agent:

forge agent \
  "Create a basic CMake project and build it"

Interactive chat:

forge chat

Review current changes:

forge review

Model Management

forge model list
forge model get

forge model use ollama qwen2.5-coder:7b
forge model use gemini gemini-2.5-flash

Tracking

Model usage:

forge usage SESSION_ID

Tool execution history:

forge tools SESSION_ID

Safety

Forge requires confirmation before executing potentially destructive actions.

Protected tools:

• write_file
• edit_file
• replace_in_file
• run_command

Example:

Tool approval required: write_file
Approve? (y/N):

Enable or disable approval mode:

forge approval true
forge approval false

Configuration

Forge stores its state under:

~/.forge/

Example:

provider = "ollama"
default_model = "qwen2.5-coder:7b"
approval_mode = true

Next

1. Context size telemetry
2. Session summaries
3. Context builder

Future

• Read-only review agent
• Project memory embeddings
• Multi-project awareness

NTH

• MCP definition

License

MIT