CozyTerm

The AI coding agent that gets better the more you use it.

CozyTerm is an open-source terminal coding assistant with a multi-model architecture, LSP integration, and a self-improving training pipeline called Forge. Every session trains your local models. Your code makes them smarter.

Built from scratch — no third-party agent frameworks.

cozyterm.com

Install

curl -fsSL https://cozyterm.com/install.sh | bash

Or with Homebrew:

brew install engindearing/tap/cozyterm

The installer handles Bun, clones the repo, and puts cozy in your PATH. Then pull some models:

ollama pull qwen2.5:7b-instruct   # orchestrator (tool calling)
ollama pull llama3.2               # chat

Usage

cozy                          # interactive TUI
cozy "fix the login bug"      # one-shot mode
cozy --plan                   # read-only analysis (no file changes)
cozy /review                  # run a custom command
cozy models                   # check model availability

Engine (always-on brain)

The engine runs in the background — watching repos, running tasks autonomously, creating PRs for you to review, and sending notifications. Every action feeds training data back to Forge.

cozy engine install           # install as launchd service
cozy engine start             # start the daemon
cozy engine status            # check if it's running
cozy engine logs              # tail engine output
cozy engine stop              # stop the daemon

Multi-Model Architecture

Each task is routed to the right specialist:

Role	Default Model	Purpose
Orchestrator	qwen2.5:7b-instruct	Tool calling and agent loop
Coder	engie-coder:latest	Code generation (Forge-trained)
Reasoner	glm-4.7-flash	Analysis, planning, debugging
Chat	llama3.2	Conversation, quick answers

All models run locally via Ollama. Swap any model in .cozyterm.json. Falls back to Anthropic API when Ollama is unavailable.

Tools

The agent has 8 built-in tools:

Tool	What it does
read	Read files with line numbers
write	Create or overwrite files
edit	Find-and-replace with uniqueness check
bash	Run shell commands (with safety blocks)
glob	Find files by pattern
grep	Search file contents
undo	Revert file changes
diagnostics	Check LSP errors after edits

Plus any tools discovered via MCP servers defined in your config.

Forge (self-improving models)

Every coding session is captured as training data. Forge fine-tunes your models and deploys better versions automatically.

cozy forge status             # training stats and collector data
cozy forge train              # run full pipeline: prepare → train → deploy → eval
cozy forge eval               # benchmark against previous versions
cozy forge versions           # model version history
cozy forge rollback           # revert to previous version

The pipeline: Collect → Prepare → Train (LoRA) → Deploy (quantize + register in Ollama) → Eval (benchmark).

Features

• LSP integration — auto-detects TypeScript, Python, Go, Rust. Self-corrects after edits.
• Permission system — interactive prompts for risky tools. Allow once, always, or deny forever.
• Custom commands — markdown files in .cozyterm/commands/. Run with /review, /test, etc.
• Session persistence — SQLite-backed conversation history with file edit snapshots.
• MCP support — connect external tools via Model Context Protocol.
• Themes — cozyterm (warm amber), dracula, tokyonight. Switch with --theme.

Project Structure

packages/
  cli/          Interactive coding agent TUI
  engine/       Always-on brain (daemon, gateway, memory, channels)
  trainer/      Forge pipeline (collect, prepare, train, deploy, eval)
  shared/       Constants, types, theme
site/           cozyterm.com
install.sh      One-line installer

Config

Project config: .cozyterm.json in your project root. Global config: ~/.cozyterm/config.json. Custom commands: .cozyterm/commands/*.md (project or global).

License

MIT