Agent settings
Configure the launch flags of every agent CLI from one schema-driven form — model, permission mode, sandbox, extra folders, environment variables — applied at the start of each session.
Every agent CLI has its own flags, its own permission vocabulary and its own idea of what “dangerous” means. The Agent settings screen puts all of them behind one form: flags applied at the start of each CLI session.
How the form is built
The form is schema-driven. Each supported CLI declares its options — type, label, help text, allowed values — and OFM renders the matching controls. That is why the wording matches the CLI you selected rather than a lowest common denominator, and why a new provider gains a full settings page without a new screen.
Settings are global per provider: what you configure for Codex applies to every Codex session, in every project.
What every CLI has
Two fields exist for all providers, and they are the escape hatches:
| Field | Behaviour |
|---|---|
| Additional arguments | One argument per line, appended verbatim to the command line |
| Environment variables | KEY=value, one per line, injected into the CLI process |
Use them for flags OFM does not model yet, or for a variable a CLI needs (KIMI_SHELL_PATH,
a proxy setting) without polluting your system environment.
One warning. ANTHROPIC_API_KEY in the environment switches Claude Code to metered API
billing. That is exactly what Open Free Max exists to avoid — see
Subscription vs API.
Per-CLI options
Each CLI page lists its full set of flags:
- Claude Code — model, effort, permission mode, allowed and disallowed tools, MCP config, appended system prompt
- Codex — approval policy, sandbox, full-auto, web search, profile, TOML overrides
- Antigravity — model, effort, execution mode, sandbox, agent, project
- OpenCode — model, agent, server port, permissions by environment
- Mistral Vibe — model, agent profile, config file
- Kimi Code — model, plan / auto / yolo, Skills folders
Cursor Agent has no flag schema yet and starts with the CLI’s own defaults.
The dangerous options
Every CLI has one, under a different name: Skip all permissions, Bypass approvals + sandbox, Auto-approve everything, Yolo. They all do the same thing — the agent stops asking.
The consequence is easy to miss: Mission Control works by detecting the questions an agent asks. An agent that never asks produces no queue entry, no rule match and no audit row. The rules engine is not bypassed, it simply has nothing to see.
If what you want is unattended work with a safety net, leave the CLI asking and let the rules engine answer the repetitive questions instead.
Language
The same screen carries the app language. Switching it re-renders the interface immediately without killing your running sessions — the terminals keep streaming while the UI changes language around them.
Frequently asked
- Are agent settings per project or global?
- Global per provider. The flags you set for a CLI apply to every session of that CLI, in every project.
- What if the flag I need is not in the form?
- Use Additional arguments — one argument per line, appended verbatim to the command line. It is the deliberate escape hatch for anything the form does not model.