Configuration

Fellow supports two complementary ways of setting configuration:

1. YAML File (recommended)
   Pass a .yml config via the --config flag:

   fellow --config my_task.yml
   

2. Command-Line Flags
   All options can also be set directly via CLI arguments. This is useful for scripting or temporary overrides.
   CLI args take precedence over YAML.

Examples:

fellow --task "Write a README.md for this project" --log.active=true --steps_limit=5

---

Default Configuration

You can view the full default configuration here:

default_fellow_config.yml on GitHub

This file contains all supported configuration fields with reasonable defaults. You can copy and adapt it for your own use.

---

Explanation of Fields

introduction_prompt

The system prompt shown to the AI. You can include {{TASK}} to dynamically insert your task text.

first_message

The first message sent to the AI after the system prompt — usually an invitation to begin acting.

task

The main instruction Fellow will execute. Should be plain text or multiline YAML string.

log

Controls whether output is saved to a file and how it is formatted.

• log.active: Enable/disable logging
• log.spoiler: Whether to wrap logs in collapsible blocks (for Markdown)
• log.filepath: Where logs are written (e.g. fellow_log.md)

ai_client

Defines which AI backend is used (openai, gemini, or a custom client) and includes its config.

For detailed per-client config, see:

• OpenAI Client
• Gemini Client
• Custom Clients

commands

Defines which commands the AI is allowed to use, and optionally applies policies to them.

commands:
  create_file: {}
  view_file: {}
  edit_file:
    policies:
      - name: deny_if_field_in_blacklist
        config:
          fields: [ filepath ]
          blacklist:
            - ".env"
            - "secrets.*"
  run_python: {}
  list_files: {}
  list_definitions: {}
  get_code: {}
  make_plan: {}
  summarize_file: {}
  pip_install: {}

Each key is the name of a registered command.
The value can be an empty dictionary {} (to allow the command without restriction), or it can include a policies: list to define checks that must pass before the command is executed.

Policies help enforce safety, access control, or project-specific constraints.

See also:

• Built-in Commands
• Custom Commands
• Policy System

planning

Optional planning mode. If active, Fellow will first call make_plan command with the planning prompt.

steps_limit

If set, limits the number of iterations (reasoning + command cycles) in the session. If null, Fellow will run until AI decides to stop.

custom_commands_paths

List of directories to search for custom commands.

It is recommended to set this to: .fellow/commands

This allows you to keep custom command implementations organized and version-controlled in your project root. Commands placed here can be registered in your config and used just like built-in ones.

custom_clients_paths

List of directories to search for custom client implementations.

It is recommended to set this to: .fellow/clients

This allows you to define and organize custom AI clients (e.g. for local models or alternative APIs) within your project directory.
Clients in this folder can be referenced in your config using their module name.

---

If any field is omitted from the config, Fellow will use its internal defaults. You can use this YAML file to define reusable workflows, ensure reproducibility, or override behavior for different environments.