5 / 7

Execution modes

Applies to the current Howse beta

Each project has an execution mode. The default is automatic; you can change it in project settings. A change takes effect on the next run. Runs already in progress keep the old mode.

automatic (default)protected
CLI settings and toolsUses your CLI's own settings and tools, plus Howse's toolsLimited to Howse's rules and tools. Your CLI settings and hooks aren't carried over.
PermissionsClaude Code skips permission checks, and Codex runs with full access and no approval prompts. Other CLIs also run without approval prompts.Agents can write only to the project and linked folders, and Howse's metadata is protected.
Keep in mindDoesn't guarantee that agents stay inside the project or leave .howse files aloneWon't start if Howse can't confirm the required protections

In both modes:

Exclude user extensions

Agent settings and Butler settings offer Exclude user MCP servers and plugins for Claude Code and Codex. It takes effect on the next run and keeps Howse tools and project configuration, including project MCP servers. Codex also disables project hooks; Claude Code also excludes plugin MCP servers. This option does not change the project's execution mode or reduce its file permissions.

Provider-option limits

Provider execution options cannot override Howse-managed permissions, tools, or instructions. Model and Effort stay in their dedicated fields; conflicting keys are rejected when saving, and reserved keys are filtered again when a run starts.

In protected mode, Codex accepts only service_tier, model_reasoning_summary, and model_verbosity in config, plus its supported turn fields (serviceTier, summary, personality). Claude Code accepts only fastMode and alwaysThinkingEnabled in settings. Other native options are ignored with a reason in run details. These options do not expand file permissions or bypass the project's execution mode.