- Uses smaller contexts well: Keeps context focused instead of burying the model in irrelevant history.
- Compacts without stopping: Local models compact more often, so compaction runs in the background with no waiting.
- Makes tool calls reliable: Validates calls and gives clear recovery feedback when the model gets them wrong.
- Breaks doom loops: Stops repeated or runaway tool behavior before the model gets stuck.
- Prevents overthinking: Cuts off excessive reasoning and gets the model back on task.
magnitude from the directory you want the agent to work in.
Work with a project
The harness can inspect and edit files, run commands, and execute scripts in the current directory. Type@ to mention a file or directory. You can paste or drag images into the composer.
Enter bash mode with /bash or start a message with ! to run a shell command directly.
Run /init to create an AGENTS.md file for project conventions, architecture, tests, and other instructions.
Sessions
Conversations are saved automatically. Use/resume or Ctrl+R to open a recent conversation and /new to start another one.
When a conversation approaches the model context limit, Magnitude compacts older work into a summary and continues the session.
Commands and shortcuts
Type/ to browse available commands. Common commands include /models, /catalog, /hardware, /setup, /bash, /init, /resume, and /new.
Skills
Skills add reusable capabilities to the harness. Install them in the current project or globally with a compatible skill installer such asnpx skills add.
Magnitude loads skills from .claude/skills, .agents/skills, and .magnitude/skills in the current project and from the matching directories in your home directory. Project skills take priority over global skills with the same name.
Review a skill before installing it. Skills can run commands, install dependencies, open websites, and interact with external services.
Web search
SetEXA_API_KEY before starting the harness to enable web search through Exa.