← Back to Discover

Kolega Code

Terminal coding agent that keeps its state on your machine, connects to many model providers and MCP servers, and can optionally let the model script parallel sub-agents for large tasks.

Coding agentsCommand-line tools
Visit website ↗

What it does

Kolega Code is an open-source coding agent you run in your terminal. It starts in an interactive text interface for day-to-day work, or runs one-off prompts from the command line with JSON output for scripts and CI. Sessions, settings, permissions and keys are stored locally. For ordinary changes it offers a read-only Plan mode and a Build mode, tools for reading and editing code and running shell commands under permission controls, web search, and a browser agent for interactive sites. For broad jobs such as repo-wide reviews or migrations, an optional feature called Gigacode lets the model write a Python program that coordinates several sub-agents. The run is saved so it can be inspected and resumed. It is provider-agnostic. The site lists providers such as Anthropic, OpenAI, Google, Mistral, Groq, xAI, DeepSeek, OpenRouter, Ollama and local Llama, and a paid ChatGPT plan can be used by signing in instead of using an API key. MCP servers can be connected as permission-gated tools, and Zed and VS Code can host the agent over the Agent Client Protocol. It needs Python 3.11 or newer.

Capabilities

  • Plan and Build modes with permission controls for shell commands
  • Optional Gigacode mode: the model writes a multi-agent workflow that can be resumed
  • 15+ model providers, including ChatGPT sign-in and local models
  • MCP server support, including OAuth-enabled HTTP servers
  • Interactive terminal interface plus one-shot kolega-code ask for scripting
  • Editor integration with Zed and VS Code via the Agent Client Protocol

When to use it

  • Running a coding agent in the terminal while keeping sessions and keys on your own machine
  • Switching between model providers or assigning cheaper and stronger models to different roles
  • Large tasks such as a repo-wide review or migration that benefit from parallel sub-agents
  • Scripts or CI jobs that need a one-shot agent call with JSON output

Alternatives to explore

  • Claude Code