# Turn a Provider Off

> Each AI provider (Claude Code, Codex, Copilot CLI...) can be switched off on one computer.

- Area: Desktop app
- Plans: All plans
- Last checked against the product: 2026-11-06
- Web page: https://agentsroom.dev/docs/provider-disable

## What it does

Each AI provider (Claude Code, Codex, Copilot CLI...) can be switched off on one
computer. AgentsRoom then treats that provider as absent, even when its CLI is
installed and on the PATH: it never launches it, never picks it automatically,
and leaves its local files alone (configuration, accounts, session history,
credentials, hooks, usage and quotas). You keep using that CLI yourself, in your
own terminal, with your own configuration, exactly as before.

## Where to find it

Settings > AI providers, open the provider's tab, then the switch "Use
<provider> in AgentsRoom" under the provider's name. A provider that is off keeps
its tab (marked "Off") so it can be switched back on.

## How to use it

1. Open Settings > AI providers and click the provider's tab.
2. Turn off "Use <provider> in AgentsRoom".
3. If it was your default provider, the default moves to the first provider still
   on. The tab shows how many agents of your open projects run on it.

Switching it back on restores everything at once: pickers, usage polls, launches.

## Settings

- `disabledProviders` (global, machine-local, Settings > AI providers > provider
  tab > "Use <provider> in AgentsRoom"): list of provider ids that are off, e.g.
  `["claude"]`. Default `[]` (every provider on). Not synced with your account on
  purpose: it governs this computer's files and processes, and the same person may
  use that provider in AgentsRoom on another computer.

## Agent tools (MCP)

- `settings_set` with `key: "disabledProviders"` (scope global): switch providers
  on or off, e.g. `["claude"]`.
- `settings_get` / `settings_list`: read the current list.

## Providers

All providers. For Claude Code, the switch also stops what AgentsRoom does in the
background without an agent running: the file-attribution hook in
`.claude/settings.local.json`, MCP approvals, folder trust, the usage API and
keychain reads, session renewal, the hidden `/usage` terminal, quota warm-ups,
`~/.claude` scans (sessions, transcripts, project discovery, external sessions),
`~/.claude.json` reads, `~/.claude/agents` imports and the `claude --version`
probe.

## Mobile

The phone hides the providers that are off on the connected computer from every
picker (New agent, lightning sheet, provider switch). The switch itself is
desktop-only: it concerns that computer's files.

## Limits

- An existing agent set on a provider that is off does not start: its console
  says why. AgentsRoom never relaunches it on another provider by itself; choose
  another provider for the agent or switch the provider back on.
- A team step, a trigger or a ticket with no provider set, which used to run on
  Claude, runs on your default provider while Claude is off.
- Quota rules that would move a launch to a provider that is off are skipped.
- Turning Claude Code off stops new writes, it does not clean up old ones: the AgentsRoom entries already in a project's `.claude/settings.local.json` stay there until you delete them by hand (they are marked `_agentsroomHook` or `_agentsroomTeamHook`). Switching Claude Code back on writes them again at the next agent start.

## Common questions

- **Can I stop AgentsRoom from touching my Claude Code configuration?** Yes: turn
  off "Use Claude Code in AgentsRoom" in Settings > AI providers > Claude Code.
  AgentsRoom then never reads or writes `~/.claude`, `~/.claude.json` or your
  `CLAUDE_CONFIG_DIR`, and never starts `claude`.
- **I use Claude Code with my own accounts in my terminal, will AgentsRoom
  interfere?** Not while Claude Code is off in AgentsRoom: no new hook, no settings change, no usage polling, no account scan. Hook entries written before you turned it off stay in the project's file, and exit at once on a run AgentsRoom did not start.
- **What exactly does the Claude Code hook write and run?** Each time an agent starts in a project, AgentsRoom keeps entries marked `_agentsroomHook` in that folder's `.claude/settings.local.json`, next to your own hooks. Most never make the agent wait: `PostToolUse` (file edits, and the tools that start or stop background work), `UserPromptSubmit`, `PermissionRequest`, `Notification`, `Stop`, `StopFailure`, `SessionEnd`, `SubagentStart`, `SubagentStop`. Two wait for your answer: `PreToolUse` on browser and test tools, and `PermissionRequest` on `ExitPlanMode` for plan review. A team run adds a reminder on `UserPromptSubmit`, marked `_agentsroomTeamHook`, and the file also lists the AgentsRoom MCP servers Claude Code may load. Each entry runs a Python 3 script from `~/.agentsroom/hooks/`, standard library only. On a run AgentsRoom did not start, it exits at once. Otherwise it posts the event (status, file path, session id, folder) to the app on `127.0.0.1` with a local token and writes a short local log, `~/.agentsroom/hook.log`. It sends nothing to the internet. On Windows with no Python, no hook is written at all.
- **Do I need to uninstall the CLI?** No. The switch is enough, the CLI stays
  installed and usable outside the app.
- **Why does my old agent not start any more?** Its provider is off. Edit the agent
  to pick another provider, or switch the provider back on.

## Related

- [Multi-Provider Support](https://agentsroom.dev/docs/multi-provider.md): every provider AgentsRoom can launch.
- [Proxy per provider](https://agentsroom.dev/docs/provider-proxy.md): another per-provider setting of the same page.
