Skip to content

Config tab

The Config tab edits tclaude's settings file, ~/.tclaude/data/config.json, in a form. Use this page to find where a setting lives. Each section has its own address, such as /config/notifications. Bare /config opens the section you last had open in this browser tab.

Using it

  • Search settings… filters rows by label, description, key or option. Every word must match, the rail shows hits per section, and Esc clears it.
  • More… on a row shows the full explanation.
  • Save changes… shows a line diff before writing anything. Dots on the rail mark sections with unsaved changes. Reload throws away your edits.
  • A linked node's config. A trusted peer can read and change this tab through the peer view, under its node.config.* grants, with every secret withheld. From the CLI, tclaude config get|set|clear --node <peer> does the same. See Remote Config tab.
  • Badges on a row say when a change takes effect or what it covers: agentd restart, refresh (next dashboard refresh), next launch, Claude Code (that harness only), deprecated, irreversible.

On a peer's page

A trusted peer's per-node page has a Config tab when that peer shares node.config.read with you. It edits the peer's settings through the remote Config tab API:

  • A row you may not change is read-only, and says what it needs: a grant such as node.config.agents, or Change on the node itself for the keys only the node changes (shell commands, the hub link, audit retention). The rail marks a section partial or read-only.
  • A secret shows only as set or not set. The peer never sends its value, not even under unrestricted trust. Where you may change it, Replace takes a new value and Clear removes it. The save confirmation names the secrets it sets or clears.

In a fused view

When a fused view spans several nodes, Config of at the top of the tab picks the node it edits. It starts on this node. A peer is edited as on its own page, through this node, and the header and the save confirmation name the node. Switching nodes with unsaved changes asks first. The sections read from this node itself (Installation and the remote-access setup) and the settings with no key link to the peer's own page instead.

Appearance

/config/appearance: how the dashboard looks.

Group Label Key Default
Activity bots Regular mode, Wizard mode 🧙, Slop mode 🎰 dashboard.activity_bots.* emoji; emoji; sprites
Tabs Plugins tab dashboard.always_show_plugins_tab off
Tabs Vegas music in regular mode slop.vegas_in_regular_mode off
Header & rows Horizontal scroll dashboard.hscroll_follow on
Header & rows Group quick options dashboard.group_quick_options hover
Overlays Preview background dashboard.overlay_backdrop dim; also blur, both, none
Overlays Window controls dashboard.window_controls auto (Standard: − ⇄ × on the far right); opt-in OS looks os (match the browser's OS), macos, windows, linux; also classic
Overlays Snap Assist dashboard.no_snap_assist off (Snap Assist on)
Header & rows Group descriptions (deprecated) dashboard.show_group_description off
Header & rows CI on finished PRs dashboard.show_final_pr_checks off
Header & rows Agent hide button dashboard.show_agent_hide_button off
Header & rows Toolbar follows the clicked group dashboard.fused_toolbar_fixed_node on (key is false)
Header & rows Collapse empty Usage views dashboard.usage_collapse_subviews off
Header & rows Open PRs indicator dashboard.always_show_open_prs, dashboard.recent_pr_window_days on; 3 days
Header & rows Side panel at the edge dashboard.side_panel_edge_reveal, dashboard.side_panel_predictive_peek, dashboard.side_panel_growing_peek on, on, off (only a change is written)
Header & rows Side panel at the edge › Advanced dashboard.side_panel_edge_zone_px, _retract_ms, _predict_min_speed, _predict_speed_distance_px, _predict_min_speed_floor, _predict_min_readings, _min_travel_px, _brake_percent, _peek_min_px 32 px, 500 ms, 1200 px/s, 600 px, 600 px/s, 4, 80 px, 70 %, 48 px; out-of-range values are clamped
Slop mode 🎰 Hide the pull lever slop.hide_pull_lever off

Terminals & windows

/config/terminals: which terminal opens, and how windows and web terminals behave.

Group Label Key Default
Terminal Terminal terminal auto-detect
Terminal Default terminal dashboard.default_terminal web
Terminal Directory picker dashboard.default_directory_picker web (remote dashboards always use it)
Terminal TUI color scheme tui.color_scheme default
Window focus Window title focus.window_title on
Window focus Raise only focus.raise_only off
Window focus Auto-tile focus.tile.enabled off
Window focus Tiling layout focus.tile.layout, .resize, .gap, .margin grid; off; 8; 0
Terminals tab Stay after last close dashboard.stay_on_terminals_after_last_close off
Web terminal attachment Resize strategy dashboard.terminal_attach.mode repair
Web terminal attachment Wait times dashboard.terminal_attach.*_ms 0; 250; 250
Web terminal tabs Auto-close tabs dashboard.close_terminal_tabs Shell tabs close when their session ends

Keyboard

Lists every dashboard shortcut by where it works, with its keys on Windows and Linux and on macOS (this browser's column is marked), and whether it still works while a web terminal has focus. See Keyboard shortcuts for the defaults.

To rebind a shortcut, click Change on its row and press the new keys; Escape cancels. The keys are recorded as pressed, so Ctrl stays Ctrl and ⌘ stays ⌘. Before you use them, the list shows what they would collide with:

  • another dashboard shortcut that can be live at the same time,
  • a browser or OS chord the page never receives (Ctrl T, ⌘ Q and the like), which blocks the choice,
  • a browser chord the dashboard would take over (Ctrl L, Ctrl F and so on), or keys a terminal program reads (Ctrl C, Ctrl R…) for a shortcut that works in terminals, which only warn.

New keys need Ctrl, Alt or ⌘ (or an F-key), so they can't take keys you type. Only tab cycling can be a single letter, digit or punctuation key. Reset puts one back and Reset all to defaults puts all of them back. Changes are saved with the rest of the form as dashboard.keybindings and apply on the next refresh of the dashboard data. Shortcuts marked fixed (the browser's own, the dialog keys, and the process editor, drawing board and viewer keys) can't be rebound yet. A peer's per-node view uses the default keys.

Voice

Spoken phrases and the opt-in hands-free settings. See Voice and dictation.

Setting Key Default
Send phrase dashboard.voice.phrases, action send on, send now
Focused phrase dashboard.voice.phrases, action wake off, hey agent; active phrases must differ
Steward phrase dashboard.voice.phrases, action steward off, hey stewie; active phrases must differ
Mark dictated messages dashboard.voice.mark_dictated on
Comma at a pause dashboard.voice.comma_pause_seconds off; 0.7 seconds when on, 0.3 to 10
Period at a pause dashboard.voice.period_pause_seconds off; 1.5 seconds when on, 0.3 to 10
Stop phrase dashboard.voice.phrases, action stop on, stop listening; ends continuous mode
Continuous mode dashboard.voice.continuous_enabled off
Continuous mode auto-stop dashboard.voice.continuous_idle_seconds 120; 30 to 3600
Wake phrase listener dashboard.voice.wake_enabled off; needs a wake phrase on too
On-device speech in this browser, not saved in config not installed
Send after silence dashboard.voice.silence_send_seconds off; 2 to 30 seconds

Notifications

/config/notifications: when and how tclaude tells you an agent needs attention. Everything still lands in Messages. See Notifications.

Group Label Key Default
Delivery Desktop notifications notifications.enabled off
Delivery Deliver via notifications.delivery os
Delivery Cooldown notifications.cooldown_seconds 5
What raises a banner Notify me when an agent… notifications.transitions idle, permission, input, error, exit
What raises a banner Human messages notifications.human_messages on
What raises a banner Access requests agent.access_request_auto_open_browser, agent.access_request_system_notification off; off
What raises a banner Presented PRs agent.present_pr_notification off
Advanced Raw transition rules notifications.transitions same list
Advanced Notification command notifications.notification_command OS default

Agents & spawning

/config/agents: joining groups, spawn guardrails, naming and default permissions.

Group Label Key Default
Terminal startup groups Auto-join session.auto_join_group on
Terminal startup groups Auto-create session.auto_join_or_create_group off
Spawn guardrails Spawn restriction agent.spawn_group_restriction on
Spawn guardrails Spawn allowed groups agent.spawn_allowed_groups empty
Spawn guardrails Spawn rate limit agent.spawn_max_per_hour 10 (0 is unlimited)
Spawn guardrails Clone cooldown agent.clone_cooldown 1m
Naming Normalize spawn names agent.spawn_name_normalize on
Naming Name the tmux session agent.spawn_label_from_name off
Permissions Default permissions agent.default_permissions empty

The two terminal startup keys and agent.spawn_label_from_name can also be read and changed from the CLI: tclaude config get, tclaude config set <key> true|false|default and tclaude config clear <key>.

Launch

/config/launch: the relaunch policy, what a resume or reincarnation does with each launch input. It is not part of config.json: each change saves at once, with no Save step, and applies on each agent's next relaunch.

The Relaunch policy table lists every launch input with the reason it behaves as it does. Group environment, group remote-control policy and sandbox profile contents have a Re-resolve | Reuse switch, and the harness-config posture has a three-way choice. Every other row shows its rule as a pill: Copied at spawn, Fixed, Live, Re-read or Spawn only. Reset to defaults puts every row back to its default. Changing the table needs relaunch-policy.manage. Without it, and always in a peer node's view, the table is read-only. What each mode means is in Launch resolution.

Context & sessions

/config/context: compaction, nudges, resume prompts and transcript retention.

Group Label Key Default
Compaction Pre-compact guard pre_compact_guard.enabled, .block_manual, .thresholds off; off; 150k of 200k, 800k of 1M
Compaction Context nudge agent.context_nudge.enabled, .min_pct, .interval_pct off; 30; 10
Sessions Resume-from-summary prompt claude_resume.threshold_minutes, .token_threshold Claude Code's 70; 100000
Sessions Transcript retention claude_cleanup_period_days Claude Code's 30; written to ~/.claude/settings.json

Usage & costs

/config/costs: cost display, estimates and rate limits. Open usage limits & budgets… opens the limits manager. See Usage.

Group Label Key Default
Cost display Cost display multipliers cost.estimate_factor, cost.harness_factors.* 1; blank inherits
Cost display Show cost on subscription cost.show_on_subscription off
Cost display Gemini billing cost.gemini_billing auto
Cost display OpenCode legacy long-context cutoff opencode.legacy_long_context_pricing_cutoff 272000
Subscription usage Poll Anthropic usage API usage.poll_anthropic_api off
Subscription usage Idle timeout usage.idle_timeout 72h
Rate limits Rate-limit gating ratelimit off
Rate limits 5-hour window, 7-day window ratelimit.five_hour_percent_max_used, ratelimit.seven_day_percent_max_used 99; 99.9 when on

Ask & scribe

/config/assistants: defaults for tclaude ask and for scribe agents.

Group Label Key Default
tclaude ask Ask — profile ask.profile none
tclaude ask Ask — sandbox ask.sandbox global
tclaude ask Ask — model ask.model sonnet
tclaude ask Ask — effort ask.effort medium
Scribes Scribe — profile scribe.profile harness default

Model and effort are disabled while a profile is selected. For Ask — sandbox options, see Ask sandbox policy.

Steward

/config/steward: the Steward. See What the Steward may do.

Group Label Key Default
Stewards Stewards steward.disabled off (stewards on); see Opting out of stewards
Access Reads steward.reads operator
Access Writes steward.writes messaging
Access Default permissions stored permissions, saved from their own editor Reads and Writes only
Launch Spawn profile steward.profile global default profile
Launch Start directory steward.dir (Browse… opens the directory picker) ~/.tclaude/scribe; a set directory must exist when the Steward is summoned
Knowledge base Source download steward.source_download ask
Name and startup Name steward.name Steward
Name and startup Startup prompt steward.startup_prompt empty

Git & forges

/config/git: what agents may reach through the credential proxy, and the keys and tokens it uses. Every key starts with agent.git_proxy., left out below. Changes apply to the next proxy call, with no restart. Open MRs and merge request badges follow within 30 seconds.

Group Label Key Default
Proxy access Allowed remotes allowed_remotes empty: the proxy is off
Proxy access Protected branches protected_refs Default: main, master
Proxy access Force push with lease allow_force_push off
Credentials SSH key ssh_key empty: the daemon's ssh-agent
Credentials GitHub token file github_token_file empty: gh's own login
GitLab GitLab instances gitlab_hosts none: gitlab.com only
GitLab GitLab token file gitlab_token_file none: the GitLab proxy refuses
  • Protected branches has three choices because the file has three meanings: Default leaves the key out, None saves [] (protection off), and Custom saves the list. A Custom list left empty is refused, so it never saves as "protect nothing".
  • GitLab instances are rows of remote host → https base URL. Listing any instance turns off the built-in gitlab.com; list it too to keep it.
  • The paths are saved as typed, and ~ expands. The form never reads the files. The save dialog's diff masks the three paths, as it masks every key named like a key or token.
  • Save checks the block: an allow-list entry that is a URL or has an empty segment, a blank protected branch, and a GitLab base that is not an https origin are named, row by row, before anything is written. Values typed into config.json by hand are not refused on load: the daemon ignores a bad entry, with a log warning where it drops one. Until such a value is corrected, though, the Config tab cannot save any section, because a save checks the whole file.

Daemon & network

/config/daemon: how agentd starts, where the dashboard listens, and remote access. Rows with the agentd restart badge need one.

Group Label Key Default
Startup Auto-launch dashboard agent.auto_launch_dashboard off
Startup System tray agent.disable_tray tray shown
Local dashboard Dashboard port agent.dashboard_port random free port
Local dashboard Dashboard bind agent.dashboard_bind 127.0.0.1
Security Persist operator token agent.persist_operator_token, agent.persist_operator_token_keychain off; off
Security Broker limits broker.enforce_limits off
Remote access Remote dashboard remote_access.enabled off
Remote access Listen interface, HTTPS port remote_access.bind 0.0.0.0:8443

Certificate management is below the Remote access group. See Operating remotely.

Installation

/config/installation: how tclaude is installed, started and updated on this machine. Its parts are stacked on one page: Start automatically, App launcher, then Updates. Each opens with a status line, tagged On, Off or Needs attention, and its main action. A part that is set up and healthy folds its detail away; Show details opens it, and a settings search shows it anyway. /config/service and /config/updates, the sections these parts used to be, open Installation at that part.

Start automatically

How agentd was started on this machine, what it should be started with, and starting it when you log in. The part reads the running daemon, and its buttons act at once, never through Save changes…. The CLI is tclaude service status|install|uninstall and tclaude service launch.

The status line says whether agentd starts on its own: not at all (it stops when you log out or reboot), from a unit tclaude wrote, or from a unit tclaude did not write. Start on login… or Remove… is next to it. It needs attention when the unit is not enabled, or runs a different command from the running daemon.

  • Launcher shows what started agentd (a systemd user or system unit, launchd, a tmux pane, a shell, or a detached process with no terminal, such as one the app launcher started), the command and environment it runs with (secret-looking values hidden), and the unit's restart policy. Restart agentd… goes through that supervisor (systemctl --user restart, launchctl kickstart -k), so the unit's own arguments stay; without one, agentd re-runs itself with the same arguments. On macOS with the tray icon showing, it starts again in a new process instead, because the icon does not come back after a re-run in the same process: a detached daemon is started detached again, and a daemon in a shell or tmux pane stays the foreground job there, so Ctrl+C still stops it. There a small helper process keeps the job and starts each new agentd, so ps shows one helper however often agentd restarts, and the old icon leaves the menu bar.

A restart behaves like a fresh start. Without a supervisor, and under launchd, the restarted agentd asks your login shell for its environment again ($SHELL -l -i, as the app launcher does), instead of keeping the one it started with. A toolchain or profile change since then (a new Go, a new PATH entry) reaches agentd, and the agents and panes it starts after the restart. What belongs to the launch itself stays: the launch profile's and the unit's variables, the TCLAUDE_* ones, and those of the terminal, tmux pane or SSH session agentd runs in. The old PATH stays on the end as a fallback. If the shell gives no answer within 30 seconds, or fails, agentd restarts with its old environment. output.log has a warning line, and a notice in Messages says so. A systemd unit is the exception: a fresh start of a unit uses the unit's environment, never your login shell, so every restart of it keeps that (see systemd). Agents already running keep their environment until they are restarted. - Launch profile is what agentd should be started with: serve flags, and environment variables, each either set to a value or inherited from agentd's own environment when the unit is written. It is kept in config.json as service.launch, but only through this section or the CLI. - Use what's running now fills the form from the running daemon. A secret-looking flag is left out, with a note: pass the secret through an inherited variable instead. A secret-looking variable is inherited, and its value is never stored. - An inherited name that agentd does not have now is flagged under its row, because the unit would get it empty. - Save profile sends the profile. A refused entry is marked on its row. If the profile changed elsewhere meanwhile, your edits are kept, and Save replaces it. Clear profile… removes it. - The section compares the running daemon with the profile, or, for a unit tclaude did not write, with that unit. The tag says which. The differences table shows values only for variables that are not secret, and a peer with restricted access sees names only. A viewer who is shown names only cannot edit the profile, so its values are never saved back empty. tclaude never changes a unit it did not write, so a profile is not applied to one. - While a profile is set, Start on login writes the unit from it, and an update compares the restarted daemon against it. - Start on login writes a systemd user unit (Linux) or a LaunchAgent (macOS) for agentd, with KillMode=process so tmux sessions survive restarts. With a launch profile set, the unit gets the profile's flags and environment. Without one, it keeps the serve flags and environment agentd runs with now by default, and use these takes your own serve flags instead, which the daemon checks one by one. The section compares the unit's command with the running one. Installing takes effect at the next login or boot. Remove… deletes only a unit tclaude wrote, and a unit you wrote yourself is shown read-only. - On Linux, agentd stops at logout unless lingering is on. That needs sudo, so the section shows the loginctl enable-linger command to run yourself.

App launcher

The app launcher adds tclaude to the desktop's app menu. Picking it opens the dashboard. The part's status line says whether it is installed, out of date, or made by hand, next to Add to app menu… (Add to Applications… on macOS), Reinstall… or Remove…; each confirm names the files written or removed. Below it, the launcher's files, the binary it runs and Start agentd if it is not running, then the serve flags: one control per default flag (and No tray icon), each saying what it does, the other flags as text, and the command they make. Rewrite launcher… applies them, and Back to tclaude's defaults… clears the launcher's own setting. When the launch profile is set, the part says it decides the flags and shows them read-only; change them in the launch profile. The CLI is tclaude service launcher status|install|uninstall (--no-start-daemon, --node <peer>), and the API is GET /api/node/launcher and POST /api/node/launcher/install|uninstall.

  • Linux: it writes a desktop entry, $XDG_DATA_HOME/applications/tclaude.desktop (default ~/.local/share/applications), its icon under icons/hicolor/scalable/apps/, and a launcher script in $XDG_DATA_HOME/tclaude/.
  • macOS: it writes ~/Applications/tclaude.app and registers it with Launch Services, so Spotlight and Launchpad find it straight away. The bundle holds a shell script and is not signed. It is marked as native on Apple Silicon (LSArchitecturePriority arm64 first), so macOS does not offer Rosetta for it.
  • The launcher runs the tclaude binary that agentd runs, with agentd's PATH. When that binary moves, or an update changes it, agentd rewrites the launcher the next time it starts. The status shows a launcher that is out of date. On macOS, Homebrew's /opt/homebrew/bin and /usr/local/bin are added to that PATH when they exist. An install warns when tmux is not on the PATH, because an agentd without tmux cannot reach any agent. agentd itself logs an error at start when it cannot find tmux.
  • Start agentd if it is not running (on by default) starts it first:
  • through the autostart unit, when Start on login installed one;
  • otherwise (or when that unit does not start) as tclaude agentd serve, logging to ~/.tclaude/data/launcher-agentd.log.

An app starts with the desktop's bare environment. On macOS that means a short PATH, no locale, and none of your shell profile's variables (API keys, proxies). So before starting agentd itself, the launcher asks your login shell for its environment. It runs $SHELL -l -i (or your account's shell), which reads the same profile files a terminal does. agentd, and every agent it starts, then gets what a terminal would give it. The launcher's own PATH stays on the end as a fallback, and the launch profile's variables still win. The launcher waits only until the shell answers, and at most 30 seconds. If the shell gives no answer by then, or fails, agentd still starts, with the launcher's environment, so the dashboard can tell you. A desktop notification says so. A notice in the dashboard's Messages says agents may not start or may show offline, and how to check (time $SHELL -l -i -c true). launcher-agentd.log has an error line. On macOS, when the autostart unit (launchd) starts agentd, agentd does this step itself (see launchd); a Linux systemd unit keeps its own environment. On macOS the locale usually comes from the terminal app rather than your profile. When neither the launcher nor the login shell names a UTF-8 locale, agentd gets your own. On macOS that is your system locale (defaults read -g AppleLocale, for example en_GB, giving en_GB.UTF-8). On Linux it is LANG from /etc/locale.conf or /etc/default/locale. If neither is set, it falls back to en_US.UTF-8 on macOS and C.UTF-8 on Linux, and only to a locale that is installed. agentd sets LANG when it has none, else LC_CTYPE.

The launcher waits for agentd to answer before it opens the dashboard. If agentd does not start, or the dashboard does not open, a desktop notification says why. - Serve flags. These are the flags the launcher starts agentd serve with. They come from the first of these that is set: 1. the launch profile's flags (the profile overrides the rest); 2. the launcher's own setting, service.launcher.serve_args (CLI: --serve-args "..."; --default-serve-args goes back to the defaults); 3. tclaude's defaults: - --persist-operator-token: the dashboard sign-in survives restarts; - --dashboard-port=<port>: a fixed port, so bookmarks and open tabs keep working. It is the running daemon's fixed port, or 7878; - --no-print-human-token: a launcher has no terminal, so the token would only end up in the log.

The tray stays on, and the dashboard stays on loopback. The status names the source of the flags in effect. The flags are written into the launcher, and a change to the profile or the setting rewrites it. A peer without full access may set only --persist-operator-token, --dashboard-port, --no-print-human-token and --no-tray. - Signing in. The launcher opens the dashboard as you. To prove that, it needs the operator token, and it has no TCLAUDE_HUMAN_TOKEN in its environment, so it reads the persisted token instead: the 0600 ~/.tclaude/data/operator_token file, or the OS keychain under --persist-operator-token-keychain. - The running agentd keeps its token in memory only: every click is refused as "unconfirmed caller". The status warns about this. Restart agentd with --persist-operator-token, or set agent.persist_operator_token. - The launcher's own serve flags do not persist the token: the status warns as well, because a daemon the launcher starts would have the same problem. - Remove… deletes only a launcher tclaude wrote. A launcher you made yourself at the same path is reported and left alone. - Peers with node.update.settings can install and remove the launcher on this node. With node.read they can see it.

Updates

How tclaude updates itself (updates.self). Everything is opt-in: nothing updates on its own until you choose a method and turn on a schedule or start-up updates. The CLI is tclaude update config show|set.

The status line says whether tclaude updates itself, with what and when. Its button picks a method when none is chosen, turns on a daily update when a method is chosen but nothing runs, and Turn off stops scheduled and start-up updates while keeping the method. These buttons edit the form, so they are saved with Save changes… like the rows below.

Group Label Key Default
Method Update method updates.self.method not chosen
Method Checkout, Branch or tag (Build from a checkout) updates.self.source_checkout.path, .ref none; the release tag
Method Update command, Timeout (Custom command) updates.self.custom.command, .timeout_seconds none; 1800 s
Method Package manager acknowledgement, Install to updates.self.package_manager_ack, updates.self.install_dir off; ~/.local/bin
When Scheduled updates updates.self.schedule.frequency off
When Day, Time window updates.self.schedule.weekday, .window_start, .window_end Monday; 03:00–05:00
When Wait while agents are working updates.self.schedule.defer_while_agents_working off
When Update when the daemon starts updates.self.update_on_start off
After an update Run setup with the new version, Setup options updates.self.post_update_setup.enabled, .flags, .harness off

The part shows how this node's tclaude was installed, with the evidence. When a package manager (Homebrew, Nix, mise, asdf, apt or rpm) owns it, the part shows the command to update it yourself. Notify me is the safe choice there. A Let tclaude run … method asks you to confirm before it is picked. A method that installs the binary itself (release, go install, checkout, custom) needs you to acknowledge that the package manager still owns tclaude. That method installs to Install to, and the part warns when that directory comes after the package manager's copy on PATH. Methods that cannot run here are greyed out with the reason.

Logs & housekeeping

/config/logs: log verbosity and rotation, and cleanup of retired agents.

Group Label Key Default
Logging Log level log_level info
Logging Startup timings startup_timing off
Logging Record hooks record_hooks off
Log rotation Max log size log_rotation.max_size 10 MiB (0 disables)
Log rotation Keep log_rotation.keep 5
Retired-agent cleanup Retired-agent auto-cleanup agent.retired_cleanup.enabled, .after_days off; 365 days

Auto-cleanup is irreversible.

Processes

/config/processes: policy for Processes.

Label Key Default What it does
Refuse unsigned workflows processes.refuse_unsigned_imports off Refuses to import a workflow file without a valid signature

Experimental

/config/experimental: feature flags, under Feature flags.

Label Key Default What it adds
Groups Route map features.groups_route_map off Group routes map view
Group attachments features.group_attachments off A group's link or attachment panel
Command palette in terminals features.terminal_command_palette_shortcut off Ctrl K (⌘ K on a Mac) inside a terminal
Recorded sandbox details features.recorded_sandbox_details off A › details link on the sandbox badge
Agent dirs: mount parent features.agent_dirs_mount_parent on See Sandboxing

Advanced

/config/advanced: Sudo (raw JSON) edits agent.sudo.

Settings with no form

These keys only exist in config.json. The form keeps them when it saves.

  • Top level: status_snapshot, host, experimental_harnesses, conv_watch, session_watch, audit, federation, updates, snapshots, routes, taskboard (Task boards retention, and the idle check's global defaults idle_check and idle_grace).
  • features: taskboard (on; false turns it off, see Task boards).
  • agent: branch_history_pr_enrichment, resource_delegation_dir, morph, spawn_legacy_injection, spawn_inline_max_chars, message_inline_max_chars, linear_proxy, awb_proxy, http_proxies, model_proxy_disabled.
  • session: tmux_name_style, auto_name_from_prompt.
  • notifications: known_types.