Project_Detail
All_Projects

Roze

Sarah's personal coding agent for use with local llms running on ollama

View on GitHub
README.md

Roze

Roze is a local coding agent built with Go, Bubble Tea, and Charmbracelet for Ollama models.

It is designed as a terminal-first assistant that can:

  • discover locally installed Ollama models
  • chat through a Bubble Tea TUI
  • inspect the workspace with list_files and read_file
  • create or replace files with write_file
  • run shell commands with explicit approval

Requirements

  • Go 1.24.2+
  • Ollama running locally
  • at least one pulled model, for example:
ollama pull qwen2.5-coder:7b

Run

go run .

If Ollama is not on the default host, set:

export OLLAMA_HOST=http://127.0.0.1:11434

Controls

  • enter: submit prompt
  • y / n: approve or deny a pending tool action
  • /model <name|index|next|prev>: switch models
  • /models: list loaded Ollama models
  • /reload: reload models from Ollama
  • /approve / /deny: approve or deny a pending tool action
  • /quit: quit

Agent behavior

Roze uses a provider-specific Ollama client and a JSON-based agent loop instead of model-native tool calling. That keeps it compatible with a wider range of Ollama-served models while still giving it a basic coding toolset.

The current MVP is intentionally conservative:

  • file writes require approval
  • shell commands require approval
  • all paths are constrained to the current workspace

Next directions

  • diff-aware file editing instead of full-file rewrites
  • richer toolset for search and patch application
  • streaming assistant output
  • session persistence and resumable conversations