Key Concepts

Fellow is built around a few foundational ideas that shape how it works and how you interact with it. Understanding these concepts will help you customize and extend it effectively.

Task-Driven Execution

Everything in Fellow starts with a task, defined in a YAML file. This task is not just a goal — it’s the seed for a chain of reasoning. For example:

task: |
  Generate a README for this project

This task is injected into a system prompt, which guides the AI in choosing and executing appropriate actions step-by-step.

Reasoning Loop

Fellow follows a simple but powerful loop:

  1. Generate a message based on the task and current context
  2. Optionally call a function (e.g. edit_file, get_code)
  3. Execute function.
  4. Log results and feed them back into the loop
  5. Repeat until the AI outputs END or a step limit is reached

This enables autonomous workflows without needing to script each step.

Plugin System: Commands

Commands are Python functions that the AI can call. Examples include:

You can define your own custom commands or override existing ones. The AI sees these as callable tools and chooses when to use them.

Learn more in the documentation:
Custom Commands

Plugin System: Clients

Fellow is not tied to a single AI provider. It uses clients that wrap LLM APIs.

Built-in clients:

You can implement your own by creating a new client class. Learn more in the documentation: Custom Clients

Memory

Fellow can persist a log of thoughts and actions between runs. This log (typically stored in memory.json) allows the assistant to remember what it did in earlier steps and avoid redundant actions.