cockpit settings

このReference本文は、アプリが配布するcockpit CLIヘルプと同じ英語の正本から自動生成されています。日本語の概念説明と安全な使い方はcockpit CLIを参照してください。

Read and change allowlisted user-facing Settings. Every command returns JSON, changes one exact key at a time, and is local-only; --host returns host_not_supported. Remote Access configuration belongs to cockpit remote-access.

cockpit settings list [--section <section>] [--changed]
cockpit settings get <key>
cockpit settings describe [<key>] [--section <section>]
cockpit settings set <key> <value>
cockpit settings set <key> --stdin
cockpit settings set <key> --file <path>
cockpit settings reset <key>
cockpit settings preview-sound ask
cockpit settings set agents.credential.<name> --stdin
cockpit settings set agents.credential.<name> --key-file <path>

Sections are appearance, notifications, shortcuts, agents, smartRouting, taskList, and onboarding. list returns effective values; --changed returns non-default values. describe includes value types, allowed values, defaults, access, and dynamic <model-id> patterns. reset affects only one key. Use none to unbind a shortcut.

set and reset return the key re-read after the write, so the reported value is what was persisted. set <key> --stdin and set <key> --file <path> read the value from standard input or a file and send it unchanged, including line breaks; use them for multi-line text such as smartRouting.policy. Empty input is rejected; use reset to clear a value. Booleans accept only true or false. Integer keys clamp to their documented range instead of failing — appearance.chatFontSizePx and appearance.pwaFontSizePx to 11–18 and taskList.widthPx to 220–520 — so check the returned value rather than assuming the requested one. shortcuts.chatSendKey accepts enter or modEnter. A shortcut that collides with another binding is rejected with shortcut_conflict, and a Quick Task accelerator the operating system refuses is rolled back with shortcut_registration_failed.

Writable appearance keys include appearance.language, appearance.mode, appearance.zoomPercent, appearance.chatFontSizePx, appearance.pwaFontSizePx, and terminal themes for Claude, Codex, Antigravity, Cursor, Qoder, Grok, and Muse Code. Dark mode is cockpit settings set appearance.mode dark. appearance.pwaFontSizePx (default 13) sets the base text size of the remote PWA on phones and tablets; the PWA task list derives its column count from that size and its width, and the same value is editable from the desktop Settings → appearance and from the PWA server sheet.

taskList.deviceFilter chooses whose tasks the desktop task list shows: local (the default) shows only this device, all adds every other device, and a device's host from cockpit devices shows only that device. Other devices are Tailscale peers owned by the same Tailscale user whose Cockpit has Remote Access on; they need no pairing and are looked up again every 30 seconds while the filter is not local, and no connection is made while it is local. Their tasks appear under the device name with status (including a pending Ask), agent, and folder only, follow taskList.sortOrder and taskList.agentFilter, and are view-only: they cannot be opened, answered, approved, completed, or deleted from this device. A selected device that is offline stays selected and is shown as offline until it reconnects. A device whose Cockpit refuses the connection, for example an older version or a different Tailscale owner, is shown as unavailable and checked again every 10 minutes; any value without spaces up to 253 characters is accepted.

cockpit settings set taskList.deviceFilter all
cockpit settings set taskList.deviceFilter studio.tailnet-name.ts.net
cockpit settings reset taskList.deviceFilter

shortcuts.followUpBehavior controls the default send action while an agent is running: queue waits until the current turn finishes, and steer sends the follow-up to the current turn. The default is queue. This does not change the separate shortcuts.steerTurn key binding.

cockpit settings get shortcuts.followUpBehavior
cockpit settings set shortcuts.followUpBehavior steer
cockpit settings set shortcuts.followUpBehavior queue
cockpit settings reset shortcuts.followUpBehavior

smartRouting.enabled (default false) and smartRouting.policy (default empty, at most 8000 characters) are the Smart routing switch and the routing policy of the new-task screen. They are the same stored values the screen edits, kept per AGI account: they belong to the account currently signed in to Cockpit, a change from either side shows up on the other (an open new-task screen updates immediately), and signing in to another account switches to that account's values. While signed out, reads return the defaults and set or reset returns sign_in_required. Right after Cockpit signs in, until it has loaded the AGI account, reading or changing these keys returns account_pending and settings list omits them; retry once sign-in finishes. Changes from the screen and the CLI are merged per key, so changing one key never overwrites the other. smartRouting.enabled only turns the switch on the new-task screen on or off; cockpit task create --smart-routing and cockpit task run --smart-routing always route with the registered policy (see cockpit help task). reset smartRouting.policy clears the policy.

cockpit settings list --section smartRouting
cockpit settings get smartRouting.policy
cat routing-policy.md | cockpit settings set smartRouting.policy --stdin
cockpit settings set smartRouting.policy --file routing-policy.md
cockpit settings set smartRouting.enabled true
cockpit settings reset smartRouting.policy

Writable agent keys include the default agent, agent commands, approval modes, supported UI modes, discovered models, reasoning effort by model, Codex service tier by model, system-prompt presets, Claude visual authentication mode, and agents.cockpit.lmStudioBaseUrl. Approval modes accept supervised, accept-edits, and full-access; agents.approvalMode.claude, agents.approvalMode.codex, and agents.approvalMode.grok also accept auto, which hands approval to the agent's own automatic review. agents.approvalMode.muse accepts only accept-edits and full-access. An Antigravity visual session runs the CLI headlessly and cannot ask for approval, so supervised makes it read-only: it auto-denies file edits and commands until the task uses accept-edits or full-access. LM Studio accepts any HTTP or HTTPS base URL, including remote hosts. Cockpit provider is writable through agents.provider.cockpit: openrouter, opencode-go, opencode, or lmstudio. Changing it selects the provider’s default model (LM Studio starts with no model; select an available model using agents.model.cockpit). Setting the same provider preserves its model. A model write also updates the provider. Provider selection and reset require the target credential: openrouter requires openrouter, opencode-go requires opencode-go, and opencode requires opencode-zen. A missing key returns credential_missing naming the credential setting. lmstudio needs no key; configure its base URL and select a discovered model. Model writes for CLI agents are checked against the last model list that discovery returned for that agent. Cockpit keeps that list when a later discovery fails or returns nothing and across restarts, refreshes it in the background, and waits longer before retrying a discovery that keeps failing. Signing in to a different account, or signing out, discards the list for that account profile; an expired session or a cancelled sign-in keeps it. A model write fails with validation_unavailable when no list has ever been discovered for that agent, including when its CLI is not installed. For cockpit, it fails whenever live discovery returns nothing, including when a provider is unreachable. Screen fallback options do not establish model availability. Codex retains its built-in fallback behavior. agents.model.muse is the exception: Muse Code's own model list can lag behind the models the account can use, so any non-empty id is kept as entered. The response carries a model_not_listed warning when Muse Code does not list it, or model_list_unavailable when Muse Code's list could not be read.

agents.enabled.<agent> (default true for every agent: claude, codex, antigravity, cursor, qoder, grok, muse, terminal, and cockpit) turns an agent on or off, the same switch as Settings → Agents → Agents in use. A turned-off agent is hidden from the agent choices of new chats on Desktop, Quick Task, and Remote Access, from Smart Routing, from the agent choices of a branched chat, and from the agent-switch menu of open chats. Cockpit also stops discovering its models: it starts no CLI and fetches no model list for it, and a saved list is served as is without a background refresh. Chats that already use the agent keep running and keep showing it, and cockpit task create --agent <agent> still starts it. At least one installed agent must stay on, so a new chat always has a choice; turning off the last one returns last_enabled_agent. Terminal and Cockpit always count as installed. Turning off the default agent moves agents.default to the first installed agent that is still on, and setting agents.default to a turned-off agent returns agent_disabled. While the Cockpit agent is off, cockpit settings refresh-models lmstudio returns agent_disabled, Cockpit model and reasoning-effort writes return validation_unavailable, and new Cockpit tasks start without fetching model metadata. Model writes for a turned-off CLI agent are checked only against the list saved before it was turned off.

cockpit settings set agents.enabled.qoder false
cockpit settings get agents.enabled.qoder
cockpit settings reset agents.enabled.qoder

Writable notification keys are notifications.sound, notifications.askSound, notifications.os, notifications.askWindow, and the Ask relay. notifications.sound is the sound for tasks that finish or need attention. notifications.askSound (default true) plays a separate rising two-note chime once when a new Ask arrives on this device; several Asks arriving close together share one chime, and reopening, restoring, or reconnecting never replays it. It is independent of notifications.sound, notifications.os, and notifications.askWindow, so the chime still plays when Ask windows are off, and when the task stops to wait within 10 seconds after its Ask chimed, the general sound is skipped so the two do not ring together. If the chime cannot play on this device, the general sound is never skipped. A device that had notifications.sound off before this setting existed starts with notifications.askSound off too. notifications.askWindow (default true) controls whether this device opens a window for each new Ask. Setting it to false closes open Ask windows without closing their Asks; every Ask stays open, is still relayed, and is answered from the side panel Ask tab (cockpit side-panel ask), Remote Access, the CLI, Discord, or Slack. Setting it back to true reopens windows for Asks that are still open. It is independent of the relay keys. Discord uses notifications.askRelay.discord.enabled, .channelId, .guildId, .allowedUserId, .attachMedia, and .acceptAttachments; Slack mirrors them with notifications.askRelay.slack.enabled, .channelId, .teamId, .allowedUserId, .attachMedia, and .acceptAttachments. The Discord id keys accept numeric snowflakes or an empty string, the Slack id keys accept alphanumeric Slack ids or an empty string, and every write restarts the relay connections. Relay tokens are never settings; the Discord bot token and the Slack bot and app-level tokens live in the secure store and are set from Settings > Ask notifications. See cockpit help ask for cockpit ask relay status [discord|slack] and cockpit ask relay test [discord|slack].

Credential entries carry key, isDefault, value:null, redacted:true, and isSet. They are agents.credential.openrouter, agents.credential.opencode-go (the OpenCode Go key, used by opencode-go/ Cockpit models), agents.credential.opencode-zen (the OpenCode Zen key, used by opencode/ Cockpit models), agents.credential.anthropic, and agents.credential.meta; the Go and Zen keys are separate and neither one substitutes for the other. Set these credentials only with --stdin or --key-file; positional secrets and inline value flags are rejected without repeating their contents. Never type a key into a command, shell expansion, or task message; pipe it from a secure source or use a protected key file. Input is trimmed and must contain 1–65536 characters. Input errors distinguish an unknown credential, missing stdin, missing/unreadable/empty files, and oversized keys without printing file paths or key contents. settings reset agents.credential.<name> removes only that key, matching clearing the GUI field. Writes share the GUI’s encrypted storage and preserve other keys. All reads and write responses remain presence-only. The anthropic credential is for Claude CLI API-key authentication, not a Cockpit provider. The meta credential is the Meta API key Muse Code uses when agents.museAuthMode is api-key; with the default muse-login, Muse Code uses the account signed in with muse login, and Cockpit removes any META_API_KEY it inherited from its own environment before starting Muse Code, so the visual UI never uses an inherited key. The terminal UI runs Muse Code inside your login shell, so a META_API_KEY that your shell startup files export still reaches Muse Code there. agents.museAuthMode accepts muse-login or api-key. Credentials never enter persisted settings or settings broadcasts. Credential changes refresh open Settings fields and renderer caches through a secret-free notification. Saving unrelated GUI settings preserves keys changed from the CLI, including resets. Runtime authentication reads the secure store; unsaved GUI keys are used only for explicit Settings model previews after the credential fields have loaded. Quick Task and other model pickers discover models using stored keys.

cockpit settings set agents.credential.opencode-zen --key-file /path/to/protected-key
cockpit settings get agents.credential.opencode-zen
cockpit settings set agents.provider.cockpit opencode
cockpit settings reset agents.credential.opencode-zen
cockpit settings set agents.provider.cockpit lmstudio

agents.museContributorModelsEnabled (default false) shows Contributor models in Muse Code model menus. Enabling it does not change the selected model. Conversations and code sent to a Contributor model may be used for training.

cockpit settings get agents.museContributorModelsEnabled
cockpit settings set agents.museContributorModelsEnabled true
cockpit settings set agents.museContributorModelsEnabled false

Secret values, resume commands, migration flags, Remote Access storage, workspace and pin fields, and renderer-local state are not exposed.

Existing settings-related commands

cockpit workspaces [list|add|create|remove|pin|unpin]
cockpit defaultAgent [get|set <agent>]
cockpit templates [list|get|delete]
cockpit templates create --name <name> --instruction <instruction> [--variable key:label:type:required:defaultValue]
cockpit templates update <id> [--name <name>] [--instruction <instruction>] [--variable key:label:type:required:defaultValue]
cockpit devices [list|all|self|resolve|alias-set|alias-remove]
cockpit usage

cockpit workspaces returns path and isPinned, and includes Creative Studio workspaces tagged with kind and a display name. cockpit workspaces create <path> creates that folder on this machine and then registers it, which is the command-line counterpart of the New folder action in the Remote work-folder picker. It creates one folder, never a chain of missing parents: the parent must already exist and sit inside an allowed root (your home folder, any registered workspace, the parent of a registered workspace, or the master folder), and the final segment must be a plain folder name. It never overwrites or merges into an existing entry. Failures answer with a code: already_exists, permission_denied, parent_not_found, parent_not_allowed, name_separator, name_relative, name_invalid_character, name_too_long, or path_too_long. Renaming, deleting, and moving folders are deliberately not offered. After saving a device alias with cockpit devices alias-set <alias> <host>, pass --host <alias> to task, autorun, accounts, fleet, hooks, or ask list/answer/close to operate a remote host. Local-only families return host_not_supported.

Devices

cockpit devices
cockpit devices self
cockpit devices resolve <alias>
cockpit devices alias-set <alias> <host>
cockpit devices alias-remove <alias>

cockpit devices lists the devices reachable by Cockpit: this machine plus every other device that answered Cockpit's health check. It is not a list of every Tailscale device, and a device that is online in Tailscale is listed only when its Cockpit answers. cockpit devices list and cockpit devices all are the same command. A device needs no alias to be listed; aliases are only names for --host. cockpit devices self returns only this machine as self and aliases, without probing other devices.

The list returns self, devices, aliases, and diagnostics. Each device has name, host, ip, os, self, and online. os comes from the Tailscale status and is one of macOS, Windows, Linux, iOS, Android, tvOS, FreeBSD, or OpenBSD; it is null when Tailscale reports no recognized OS, so show it as unknown. A device is never dropped because its OS is missing.

An empty list of other devices is only conclusive when diagnostics is clean. diagnostics.tailscaleStatus is ok, not_running when Tailscale is installed but not connected, or unavailable when the Tailscale status could not be read; when it is unavailable, self is null and devices is empty. diagnostics.unreachable lists each probed device that did not answer, with a reason of timeout (no answer in time), unreachable (connection failed), or not_cockpit (something else answered). diagnostics.notProbed counts devices skipped because Tailscale reports them offline or without an address. Devices that answered are still returned when others fail.

Browser Identities remain under cockpit browser identity. Agent account profiles can be listed, created, removed, and signed in with cockpit accounts, and switched per task with cockpit task account <id> <account>. cockpit usage reads agent usage and cockpit usage reset spends a Codex reset credit. See cockpit help browser, cockpit help accounts, and cockpit help usage.

Preview the Ask sound

cockpit settings preview-sound ask

Plays the Ask chime once on this device, whatever notifications.askSound is set to, and returns {"sound":"ask","played":true}. It changes no setting. A device with no usable audio player returns sound_unavailable. Settings > Notifications has the same Preview button; opening Settings never plays it.

Reload LM Studio models

cockpit settings refresh-models lmstudio

Fetches the Cockpit Agent's available LM Studio models again, bypassing the model cache and returning provider and models. Uses the saved agents.cockpit.lmStudioBaseUrl, even if another provider is selected. A reachable server with no available models returns an empty models array. Failed discovery returns validation_unavailable with the LM Studio fetch error. Concurrent refreshes share a fresh request, while an older in-flight list request still completes for its original callers. This read-only action does not change the selected model, restart or unload models, or interrupt tasks. In Settings, Reload models beside the model selection uses the current LM Studio URL field, including unsaved edits. A missing selection stays visible and is flagged until you choose an available model.