Omniscio

More settings (A)

Detailed settings reference for every key beginning with the letter A, covering display options, agent controls, alarms, alternative artificial-intelligence providers and automations.

On this page

Detailed reference for every setting whose key begins with the letter A, covering display options, agent controls, alarms, alternative AI providers, automations, and more.

Accounts classic layout

Switches the Accounts settings page back to the previous flat provider list instead of the newer grouped "Other AI tools" layout. Turn this on if you prefer the simpler, single-column view of all configured providers. Off by default; the grouped layout is the current standard. No restart required.

Enable the Ad Studio ad-finding skill

An optional skill that lets your own agent search Meta's public Ad Library for long-running ads and hand them to Ad Studio's import door, where you approve them in your inbox. It works through Agent Reach in a Chrome profile signed in to a separate Facebook account made just for this and tied to nothing personal — never your own account. Meta limits automated collection of its pages, so it starts at 10 searches a day (you can change that once it is on) and stops for the day at the first sign of a block. Turning it on installs it at the next launch; off removes it. Needs Node.js 18 or newer plus Agent Reach with OpenCLI. Off by default. Find it in Settings → Features.

Agent inbox alerts

When on (the default), sessions can drop a persistent row into your inbox by calling POST /alert on the local CLI control server. This is how agents surface status updates, completion notices, or anything they want you to see without stopping to ask. Turn it off to refuse cards posted from outside the app: an agent's POST /alert answers 403 and a plugin's card is not posted, while the agents themselves keep running. The cards Omniscio raises for itself — a failed backup, an account to sign in to again — are never affected.

Update alerts for Claude Code and Codex

When on (the default), Omniscio checks about every six hours whether Claude Code or Codex CLI has fallen behind its latest release and, if so, puts one card per tool in your inbox naming the installed and latest versions. The card's Update now button runs that tool's own updater in the background on your computer — from your phone, it runs on the computer — and the card goes as the update starts, rather than staying there still asking for the version you just ordered. If the update does not take, a different card tells you so, naming the version you are still on, with a Try again button (Claude Code no longer needs you to close anything for this: an update it cannot run while your sessions are using it is parked and runs by itself once they finish); either way the card also clears by itself once the tool is current. Each new version is announced once; a card you dismiss stays quiet until a newer version appears. Omniscio never updates Claude Code or Codex on its own. Turn it off to stop new cards; you can still update either tool any time from the Toolchain tools list.

Update alerts for your other tools

When on (the default), Omniscio checks about every six hours whether any of your other recommended tools — Git, GitHub CLI, OpenCode, Pi, Portless, Agent Browser, pnpm, yt-dlp, Bun, gws, kubectl, RTK, and gog, plus Playwright, AgentMail and Repomix while Auto-update Tools is off — has fallen behind its latest release and, if so, puts one card per tool in your inbox naming the installed and latest versions. The card's Update now button runs that tool's own updater in the background on your computer, the same as the Toolchain tools list's Update button, and the card goes as the update starts; if the update does not take, a different card says so and offers Try again. RTK and gog are installed by hand, so their card says so and its button opens that tool's download page instead; the card also tells you which file to download there and to unzip it over the old program. Each new version is announced once; a card you dismiss stays quiet until a newer version appears. This switch is separate from the Update alerts for Claude Code and Codex switch above — turning either off leaves the other exactly as it was. Turn it off to stop new cards for these tools; you can still check for and update them any time from the Toolchain tools list.

Replies per day

Sets the most replies from Team Chat that Omniscio will wake your agent for in a day — ONE number for the whole day, counted per agent session across every conversation it is watching and every subscription it holds, not one allowance per conversation. When an agent posts in Team Chat and is waiting to hear back, Omniscio passes the replies to it as they arrive. Once the day's limit is reached, further replies wait, nothing is lost, and the agent is told they are waiting. They are delivered after midnight UTC, or as soon as you raise the limit. Leave the box blank for no limit. You find it in Settings > Team Chat, beside the Team Chat notification options.

Default: blank (no limit).

Max concurrent agent-driven sessions

A global cap on how many agent-driven (externally scripted) helper sessions may run at the same time. Once this count is reached, any additional spawn request is refused until one finishes. A driven session that leads a crew does not use one of these places — it uses a crew-lead place instead (next setting). Range: 1–20 — there is no "unlimited" value, so the highest you can set is 20. Applies only to sessions started by external scripts via the CLI; sessions you start yourself are not counted.

Max concurrent crew leads

How many agent-driven sessions that lead a crew (a standing overseer such as a PR Desk or a mission crew's lead) may run at once without taking a helper place. Helpers plus crew leads together can never pass this number added to the helper limit, so a helper that makes itself a crew lead still cannot raise how many paid sessions run. You find it in Settings > CLI Control, right under the helper limit. Range: 1–20. Default: 6.

Max dollars per agent-driven session

A USD ceiling for a single agent-driven session. Not enforced today: Omniscio stores this value and shows it back to you, but nothing stops a session when its cost reaches it, so treat it as a note to yourself rather than a brake. Range: 0.01–20 — there is no "no limit" value, so the loosest you can set is $20. The concurrent-sessions cap below is enforced; this one is not.

Max turns per agent-driven session

A ceiling on the number of follow-up turns an agent-driven session may take. Not enforced today: Omniscio stores this value and shows it back to you, but nothing stops a session when it reaches the count, so treat it as a note to yourself rather than a brake. Range: 1–200 — there is no "unlimited" value, so the highest you can set is 200. The concurrent-sessions cap below is enforced; this one is not.

Allow agents to ask you for a value

When on (the default), a blocked session can pop a labeled modal asking you to fill in a value it must never store itself — most often an API key. You type the value, press Enter, and it goes straight to the proper Settings field; the agent never sees what you typed, only that the request was filled or cancelled. Turn this off to refuse all agent-triggered input prompts; the /input-request route returns 403 and the session must handle the refusal on its own.

Agent instructions sync — canonical file

Which file Omniscio treats as the source of truth when syncing agent instructions to other tools. CLAUDE.md is the default. Switch to AGENTS.md if that is your primary file. The file you pick is never written as a mirror target — only the other tools get copies. Only relevant when the Agent Instructions Sync feature is on.

Agent instructions sync — hubs switched off

The hubs you have switched sync off for, one at a time, while it keeps running everywhere else. Omniscio leaves a switched-off hub's copies exactly as they are: it never writes them or its .gitignore, raises no "edited outside Omniscio" alerts for it, and shows no synced-copy rows in its Auto Context panel. Codex, Gemini and other non-Claude agents working there still get the hub's current rules in their opening instructions. Switch a hub off with the Sync this hub switch (Settings → Features → Agent Instructions Sync, for the hub you are in) or the Stop syncing this hub button in the "edited outside Omniscio" dialog. Switching it back on syncs it once straight away. Empty by default.

Agent instructions sync — git handling

Controls whether the generated mirror files are tracked in git. local (default) keeps the copies out of git to avoid merge conflicts — useful when only you use the repo. committed adds them to the repo so teammates who use other AI tools (Codex, Cursor, Copilot, etc.) pick them up automatically. Switch between the two at any time; only newly written files are affected.

Agent instructions sync — mirror targets

The list of agent-instruction filenames to keep in sync alongside your canonical file. Accepted targets: AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules, .clinerules/base.md, .github/copilot-instructions.md, .windsurfrules. Each checked target gets an up-to-date copy of your canonical file every time it changes. The canonical file itself never appears in this list.

Agents use the Agent Task System

When on, agents file their work in Omniscio's own agent task system: one list per project, separate from anything you wrote, kept up to date as the agent works. An agent's own to-do checklist stays its own and is never copied into that list. It needs the Agent Status Board on as well. This switch only decides where agents file shared work — your own tasks are never touched. Off by default. Find it in Settings → Workflow, under the Agent Status Board switch.

Which model each agent tier uses

Your preferred Claude model for each of the four agent-facing tiers — genius, smart, worker, and basic. When an agent spawns another session it asks for a tier ("worker") rather than naming a model itself, and Omniscio looks the real model up here. That is deliberate: an agent's memory of model names is frozen at its training cutoff, so a model id it half-remembers can be wrong or unlicensed and the spawn then dies. Leave a tier unset and it uses the built-in default for that tier. Setting one here changes every future spawn that asks for it. One exception to “built-in default”: an agent that asks for the worker tier without naming an engine gets the cheap DeepSeek worker model rather than a Claude model, so the many helpers an agent starts run on the low-cost lane. Pin the worker tier to a Claude model here and those helpers stay on Claude; an agent can also ask for Claude explicitly. Settings → Sessions → “Models agents may pick”.

Short share links by default

When on (the default), agents that publish a share are told to prefer the short omnisc.io link format and pass shortLink: true in their publish call. This only takes effect if your account holds the custom_share_url paid entitlement — the server enforces eligibility, so the setting is harmless on a free plan. Turn it off to stop agents defaulting to short links even on a paid plan. Your own in-app publish dialog is not affected.

Agent Screen Capture (CLI)

Lets a session ask to capture the screen over the CLI control server — a screenshot, or a recording start/stop. Off by default, and this is a security-relevant switch: while it is off the CLI capture routes answer 403 and no capture can be started by an agent at all. Turn it on and nothing is recorded yet either — each request still comes to you for approval, and only arming auto-approval records without a fresh prompt. Restart is not required. See Settings → Optional Features.

Agent Web browser

Which browser installed on this computer the agents' own browsers use, for Agent Web (in development, turned on in Settings → Lab). Chrome, or Edge if Chrome is missing (the default) picks Google Chrome and falls back to Microsoft Edge only when Chrome is not installed; Google Chrome only and Microsoft Edge only never fall back — an agent then gets a clear "no browser" answer instead. On the default, a session that already has a profile keeps the browser that profile was made by while that browser is installed, so installing Chrome later signs no agent out; only a session whose browser was uninstalled moves to the other one. Switching to the other browser explicitly starts agents on fresh browser profiles, because Chrome and Edge cannot share one, so sites they were signed in to ask them to sign in again. Your own Chrome or Edge and its profile are never touched.

Max agent browsers

How many agents can have their own Agent Web browser open at the same time (in development, turned on in Settings → Lab). Each one is a separate, invisible Chrome that uses roughly 200–300 MB of memory, so the cap keeps many busy agents from slowing your computer. When the cap is reached, the next agent waits in line for a free place — up to 45 seconds — and is then told the browsers are busy. Range: 1–5; the default is 3. A browser left idle for 15 minutes closes by itself and frees its place.

Turn on Omni

Turns on Omni, your AI project manager. Omni gives you one page in the sidebar with your projects and sessions, and one conversation that understands every project in Mission Control. While it is on, Omniscio also runs a few background helpers for it, for example one that watches your projects for work at risk. Turning it off stops them. Turning Omni on also switches on Mission Control, the PM Feed digest (a daily 8 AM "What changed" card in Omni's inbox) and the vault if you have already chosen a vault folder, whichever way it is turned on; turning Omni off leaves those three as they are.

To answer you, Omni reads your Mission Control projects: board item names, comments, documents and the people assigned to them. That content, including other people's names and comments, is sent to the AI provider behind your default API key or the model you chose. You can turn it off any time. You find it in Settings > Mission Control, and it only appears once Mission Control is on.

Default: off.

Show the Inbox tab on the Omni page

Decides whether the Omni page has its own Inbox tab. When it is on, the tab lists the notifications Omni has raised, newest first, so you can read and act on them without leaving the page. When it is off, the tab is hidden, but those notifications still arrive in your main Inbox, so nothing is lost. You find this switch in Settings > Mission Control, in the Name and look card of the Omni area. It stays there, and keeps its choice, even while Omni is off. The label shows the custom name if you gave Omni one. That feature is still being rolled out, so you may not see it yet.

Default: on.

Order of the Always here list

No longer in Settings. The Omni page used to let you choose the order of its Always here list. Every notification now lives in the Inbox tab, so there is no list left to reorder, and the saved choice has no effect.

Default: What changed first, then Suggestions.

How much Omni does on its own

Chooses how much Omni may change in Mission Control without asking you. Only suggest: every change to a project, an item or a task waits as a card under Needs your OK, apart from sharing or publishing a board and setting up an automation from your own words. Ask first: those changes wait too, except one you asked Omni for directly in that chat in the last 30 minutes. Act, then tell me: changes happen, and each item change shows in What changed with a star and Undo. At every level Omni still asks before deleting or archiving, reassigning people, or starting an AI session from an item or a task; sharing or publishing a board, and setting up an automation from your own words, happen with no card. For your Tasks the rule is looser: ordinary task changes — creating, editing, moving or completing one — happen with no card at Ask first and Act, then tell me and wait only at Only suggest, while the app still always asks before Omni deletes, archives or clears anything in your Tasks, or starts a session from a task. The app checks this itself, for changes Omni sends to Mission Control and Tasks addresses only — an address outside those two, or a call that names no session, is not checked. You find it in the gear Settings window on the Omni control center and in the How Omni answers card of the Omni area in Settings > Mission Control; both are the same setting.

Default: Ask first.

Omni page colors

Decides which color the Omni page uses for its buttons, selected tabs and focus rings. Off, the default, the page and the windows it opens use Mission Control's teal. On, they use your own Omniscio theme color instead. This is Omni's own switch: Mission Control has a separate "Match my Omniscio theme" switch for itself, and neither one changes the other. You find it in the gear Settings window at the top of the Omni control center, and in the Name and look card of the Omni area in Settings > Mission Control; both are the same switch. It stays there, and keeps its choice, even while Omni is off.

Project page order

Chooses the order of the cards on a project's page, the page that opens when you pick a project on an Omni control center: Worth knowing, Needs your OK, Last 24 hours and Project Forecast. The project's name and Omni's take always stay at the top; only these four move. Drag a card by its handle onto another row, or use its up and down arrows, which also work by keyboard and touch. Back to the default order puts them back. The phone shows the same order, and the Project Forecast shows only on a computer, so that row says Computer only. You find it in Omni's own settings window, the gear at the top of the Omni control center, under Look. It is not on the Settings page.

Default: Worth knowing, Needs your OK, Last 24 hours, Project Forecast.

Suggestions to show in Omni

Chooses which of Omni's suggestions are listed. High only shows just the most urgent findings, Medium and up adds the moderate ones, and All (the default) shows everything. A hidden suggestion is filtered out of the list, never deleted, so moving the slider back brings it straight back. You find it in Settings → Mission Control, in the Omni area, and in Omni's own control-center settings.

Alarm NLP daily cost cap

The maximum amount (in USD) that Omniscio may spend per day on AI-assisted natural-language alarm parsing — the Haiku calls that understand phrases like "remind me at 3:30 next Tuesday" instead of requiring a rigid date picker. Once the cap is reached for the day, the NLP parser pauses and only exact-format times are accepted until midnight. Range: 0–5, and the cap is checked as "spent so far ≥ cap", so 0 does not mean unlimited — it means never: at zero the cap is already reached on the first request of the day, and the parser is off entirely until you raise it. 0.1 (the default) is the tightest useful setting; raise it to spend more.

Bring app to foreground on alarm

When on, Omniscio pops its window above all other apps the moment an alarm fires, so you can't miss it. When off, the alarm notification still appears (as a modal, banner, or silent banner depending on your assertiveness setting) but the app stays wherever it is. On by default; turn it off if you would rather your current window were left alone.

Snooze cap

The maximum number of times any alarm can be snoozed before only the Dismiss button remains. Set it to -1 for unlimited snoozes. Range: -1–20. Default is unlimited (-1). Raising it to a low number (like 3) helps prevent indefinitely snoozing an alarm you keep meaning to deal with.

Default snooze length

How long a snooze lasts when you tap Snooze on a firing alarm. Accepted values: 5, 9, 15, or 30 minutes. Default is 9. You can still override the snooze length on individual alarms; this setting controls what the Snooze button does without a choice.

Default alarm sound

The sound file played when an alarm fires. Enter the filename or path of the audio file you want to use. Leave it blank to use the built-in default chime. Only applies when a per-alarm sound override is not set.

Default alarm sound duration

How many seconds the alarm sound plays before it stops automatically. Set to -1 to let the sound play in full without cutting it off. Range: -1–300. When the play mode is set to loop, this cap still applies and stops the loop after the configured duration.

Default alarm sound fade-in

How many seconds the alarm volume takes to ramp from silent to full when it starts playing — a gradual crescendo instead of a sudden blare. Set to 0 for no fade-in (the sound starts at full volume). Range: 0–300. Useful for a gentler wake-up experience.

Default alarm sound play mode

Whether the alarm sound plays once and stops, or loops continuously until you dismiss or snooze. once plays it a single time; loop keeps repeating it. Default is loop — the alarm keeps sounding until you act, stopped by the duration cap above. Choose once if you would rather hear the chime a single time.

Speak alarm label

When on, the alarm reads its own label aloud (using text-to-speech) the moment it fires, so you know what it's about without looking at the screen. Useful for accessibility or for use while away from your desk. Off by default.

Enable alarms

Master switch for the Alarms feature. When on, scheduled alarms fire at their set times with sounds, notifications, and any foreground or focus-pierce behavior you have configured. Turn it off to silence all alarms without deleting them; they resume firing when you turn it back on.

Max ringing time

How many minutes an alarm can ring before Omniscio automatically dismisses it to your inbox. Once this limit is reached the alarm stops ringing and leaves an inbox row you can review later. Range: 1–60 minutes. Prevents an alarm from ringing indefinitely if you step away.

Pierce Focus Mode

When on, alarms fire immediately even if Focus Mode is active and batching notifications. This lets urgent reminders break through without waiting for a focus-period to end. When off, alarms are held with other notifications until the current focus batch releases. On by default; turn it off if you would rather alarms respect Focus Mode's batching.

Show me the first alert of a new kind before it goes to a crew

Omniscio can send some automatic system alerts straight to a crew instead of to you. With this on, the first time a new kind of alert is raised, its card also stays in your inbox so you see it once before it becomes routine. Every later card of that kind follows your routing settings. To keep this from getting noisy, it applies to at most 20 new kinds of alert a day. If you have already set a route for a kind or muted it, your choice always wins. Turn it off if you would rather new kinds be routed from the start. You find it in Settings > Inbox.

Default: on.

Default project for sessions started from an alert

Which project a new session opens in when you start one straight from an inbox alert (and from the drip and "start a session" dialogs that share the same default). Leave it on Automatic and Omniscio picks this install's own repo hub when it has one, otherwise your Claude folder. A card that already remembers its own repo always wins over this. If the project you picked is later removed or can't host a session, the choice is quietly ignored rather than breaking the launch.

Crew that handles alerts automatically

Names the crew that acts as your alert triage desk. The triage desk is the crew whose lead receives certain automatic alerts instead of each alert starting its own fixing session. Leave this empty and Omniscio uses the built-in crew that sorts alerts. Fill it in only if you want a different crew to do that job. The crew's id is the one shown on the crew in the Overseer panel. It has no effect while "Send every automatic alert to the triage desk" is off, or while the chosen crew has no running lead. You find it in Settings > Inbox.

Default: empty, which uses the built-in alert-sorting crew.

Send every automatic alert to the triage desk

Decides what happens to an automatic alert that Omniscio would normally answer by starting a fixing session of its own. When this is on and the triage desk crew has a running lead, the alert is handed to that lead instead. The lead decides whether it is a new problem or part of one it is already fixing, so you do not end up with several sessions chasing the same fault. If no crew lead is running, nothing changes and the alert starts its own session as before. Turn it off to make every such alert start its own session again. You find it in Settings > Inbox.

Default: on.

Allow Anti-Gravity sessions in any project

Opt-in toggle that lets you start Anti-Gravity (Google's Gemini-backed coding CLI) sessions across all your projects from the new-session menu. Off by default — no Anti-Gravity sessions can be started until you turn this on and configure the Anti-Gravity CLI. Changing it takes effect immediately.

Allow Codex sessions in any project

Opt-in toggle that lets you start OpenAI Codex CLI sessions across all your projects. Off by default. You also need the Codex CLI installed and an OpenAI API key saved before sessions will actually work. Turn it on to add Codex to the provider menu.

Allow Cursor sessions in any project

Opt-in toggle that lets you start Cursor (cursor-agent CLI) sessions from the new-session menu in any project. Off by default; requires the Cursor CLI on your PATH and a Cursor API key. Turn it on to expose Cursor as a selectable provider.

Allow DeepSeek sessions in any project

Opt-in toggle to expose DeepSeek (via its Anthropic-compatible API endpoint) as a provider in the new-session menu. Off by default. You need a DeepSeek API key saved in Settings → Accounts before sessions will run. Turn it on to add DeepSeek to your provider list.

Allow Devin sessions in any project

Opt-in toggle to expose Devin (Cognition's autonomous software-engineer API) as a provider in the new-session menu across all projects. Off by default; requires a Devin API key. Turn it on to be able to start Devin sessions from within Omniscio.

Allow DeepSeek Harness sessions in any project

Lets you start DeepSeek Harness sessions from the new-session menu in any project. DeepSeek Harness is DeepSeek's own coding command-line tool, called dsh. It is separate from the DeepSeek provider, which runs DeepSeek models inside Claude Code. Until this is on, DeepSeek Harness does not appear as a choice. You also need the dsh program installed, and the same DeepSeek API key the DeepSeek provider uses saved in Settings > Accounts. This is a developer preview.

Default: off.

Allow Gemini sessions in any project

Opt-in toggle that lets you start Gemini CLI sessions across all your projects. Off by default. You also need the Gemini CLI installed (npm install -g @google/gemini-cli) and a Gemini API key configured. Turn it on to add Gemini to your new-session provider options.

Allow GLM sessions in any project

Opt-in toggle to expose GLM (Zhipu z.ai's Anthropic-compatible API) as a session provider in the new-session menu. Off by default; requires a GLM API key saved in Settings → Accounts. Turn it on to add GLM to your available providers.

Allow GPT sessions in any project

Lets you start GPT sessions from the new-session menu in any project, running on your own ChatGPT subscription. Omniscio does this through a small helper program on your computer that translates between the two, so you sign in with ChatGPT instead of pasting an API key. Your login stays on this computer. Turn this switch on first, then sign in in the "GPT, Grok & Claude" card in Settings > Accounts. Until it is on, GPT does not appear as a choice. Once you have signed in, that provider's card in the "GPT, Grok & Claude" section shows "Signed in as" with the account.

Default: off.

Allow Grok sessions in any project

Opt-in toggle to expose Grok Build (xAI's coding CLI) as a provider in the new-session menu. Off by default; requires the grok CLI installed and an xAI API key. Turn it on to start Grok sessions from within Omniscio.

Allow Hermes sessions in any project

Opt-in toggle to expose Hermes (Nous Research's local model CLI) as a provider in the new-session menu. Off by default; requires the Hermes CLI on your PATH and a model set up via hermes setup. Turn it on to add Hermes to your provider list.

Allow Kimi sessions in any project

Opt-in toggle to expose Kimi (Moonshot's Anthropic-compatible API) as a session provider in the new-session menu. Off by default; requires a Kimi API key saved in Settings → Accounts. Turn it on to add Kimi to your available providers.

Allow Kimi Code sessions in any project

Opt-in toggle to expose Kimi Code (Moonshot's own ACP coding CLI, separate from the Kimi API) as a provider in the new-session menu. Off by default; requires the kimi CLI installed and authenticated via kimi login. Turn it on to start Kimi Code sessions from within Omniscio.

Allow Meta sessions in any project

Lets you start sessions that use Meta's Muse Spark model from the new-session menu in any project. Omniscio reaches it through Meta's Anthropic-compatible API, so it runs inside the same Claude Code tool as your other sessions. Until this is on, Meta does not appear as a choice. After turning it on, add your Meta API key in the field below it in Settings > Accounts. You get the key from dev.meta.ai, which is a US preview.

Default: off.

Allow MiniMax sessions in any project

Opt-in toggle to expose MiniMax (Anthropic-compatible API) as a session provider in the new-session menu. Off by default; requires a MiniMax API key saved in Settings → Accounts. Turn it on to add MiniMax to your available providers.

Allow OpenCode sessions in any project

Opt-in toggle to expose OpenCode as a provider in the new-session menu across all projects. Off by default; requires the OpenCode CLI on your PATH. Turn it on to be able to pick OpenCode when starting a new session.

Allow OpenRouter sessions in any project

Opt-in toggle to expose OpenRouter as a provider in the new-session menu across all projects, so you can start coding sessions on OpenRouter instead of Claude. Sessions reuse the same Claude command line, redirected to OpenRouter's Anthropic-compatible endpoint — the “connect with OpenRouter, no account needed” path offered during setup. Off by default, like the other vendor spawn toggles. It is switched on for you when you connect with OpenRouter during onboarding or from Settings → Accounts, and it reuses the same OpenRouter key that powers Omniscio's built-in AI helpers — one key, no second account to set up.

Allow Pi sessions in any project

Opt-in toggle to expose Pi (Earendil's coding agent, installed via npm i -g @earendil-works/pi-coding-agent) as a provider in the new-session menu. Off by default. Turn it on to start Pi sessions from within Omniscio.

Allow Qwen sessions in any project

Opt-in toggle to expose Qwen (Alibaba's Anthropic-compatible DashScope API) as a session provider in the new-session menu. Off by default; requires a Qwen (DashScope) API key saved in Settings → Accounts. Turn it on to add Qwen to your available providers.

Allow Reason sessions in any project

Lets you start sessions with Reason, the cloud coding agent from Reason Machines, from the new-session menu in any project. Reason runs in Reason's cloud rather than on your computer, and its usage is billed to your own Reason account. Turn this on, then paste your Reason API key in the field that appears below it in Settings > Accounts. A button there opens the page where you create the key. If you switch this off later, your saved key is kept.

Default: off.

Allow T3 Code sessions in any project

Lets you start T3 Code sessions from the new-session menu in any project. T3 Code is an open-source agent control panel that runs through a program called "t3" on your computer. Until this is on, T3 Code does not appear as a choice. You also need the t3 program installed; Settings > Accounts has a "T3 Code program" row that points you to t3.codes to get it. Changing the switch takes effect right away.

Default: off.

Allow Ultramarine sessions in any project

Lets you start Ultramarine sessions from the new-session menu in any project. Ultramarine is an AI service from Paper Instruments that Omniscio reaches through the same Claude command line, pointed at their service. Until this is on, Ultramarine does not appear as a choice. You also need an Ultramarine API key, which you create at platform.paperinstruments.com and paste into the Ultramarine key field in Settings > Accounts, next to this switch.

Default: off.

Allow unverified plugin backends

Retired — there is no such setting any more. It used to be a global switch permitting any marketplace plugin to run a backend process with full system access. It is gone because a per-plugin consent record already answers that question, and a global switch answered it worse: it could overrule a plugin you had refused, and it could not answer for one you had never been shown. A plugin's backend now starts while, and only while, your own approval of that plugin is current — the version, the permissions and the fact that it ships a worker. Approve the plugin and it runs; withhold it (or let the approval fall behind a new version) and the plugin card shows Needs your permission with a Re-approve button. Built-in plugins are always allowed.

The key itself is kept in the settings schema, unread and unwritten, so an older build you downgrade to finds the value where it left it.

Allow Grok (subscription) sessions in any project

Lets you start Grok sessions from the new-session menu in any project, running on your own Grok subscription. Omniscio does this through a small helper program on your computer that translates between the two, so you sign in with Grok instead of pasting an API key. Your login stays on this computer. Turn this switch on first, then sign in in the "GPT, Grok & Claude" card in Settings > Accounts. Until it is on, this Grok choice does not appear. It is separate from "Allow Grok sessions in any project", which uses the Grok command-line tool and an xAI key. Once you have signed in, that provider's card in the "GPT, Grok & Claude" section shows "Signed in as" with the account.

Default: off.

Always use my default start config

When on, the tool/model/thinking picker that normally appears when you open a new session is hidden, and every session starts with your saved defaults immediately. You can still change the model for a single session before you send its first message. Turn it on to streamline session creation when you almost always want the same setup.

Make sessions Omniscio-aware

When on, a short note is silently added to every session's system prompt telling the agent that it is running inside Omniscio and explaining how to reach its local control server (CLI routes, alert posting, and so on). This allows capable agents to surface results in your inbox, publish shares, or request input — features they can't use if they don't know the app is there. On by default; turn it off to keep the note out of every session's system prompt.

Anti-Gravity virtual project

Adds a dedicated Anti-Gravity virtual project to the sidebar, giving you a fixed place to launch and track Anti-Gravity sessions without them mixing into your other projects. Off by default; requires the Anti-Gravity CLI and allowAntigravitySessionSpawn to be on. Turn it on once Anti-Gravity is set up to keep your Anti-Gravity work organized.

Archive digest cadence

How often Supermail automatically posts an "Archive Insights" digest to your inbox — a periodic AI summary of the senders whose emails you archive without reading, with suggestions for cleanup. Options: off, daily, or weekly (default weekly). You can always run an on-demand analysis inside Archive Insights regardless of this schedule; set it to off to stop the automatic posts entirely.

Archive Insights AI digest

Turns on Archive Insights, an AI helper in Supermail that looks at the senders whose emails you archive without reading and suggests cleanup actions for them. To make specific suggestions, the sender and subject of the emails you archive are sent to the AI provider, so leave it off if you are not comfortable with that. When it is off, the analysis does not run at all. It also needs you to be signed in. The "Automatic digest" choice that appears under it controls how often a digest is posted to your inbox. This feature is still in development, so it may not be available on your install yet. You find it in Settings > Supermail.

Default: off.

Pause before archiving a running audit

A Nighty Tidy audit is expensive and takes hours, so a stray middle-click on its session should not throw it away. When this is on, archiving a session that is still working an audit run does not happen straight away: a message says which audit is running, and Archive anyway archives the session at once. Only the session actually doing the work is held — a session that has already ended, or an older session whose work moved on to a successor, archives exactly as it does today, and so does any session with no audit running. The app's own automatic archives (a session filing itself away, cleanup sweeps) are never held. On by default; turn it off in Settings → Sessions → Behavior if you would rather every archive go straight through.

Warn before archiving loses work

Before you archive a session, Omniscio checks whether anything is still due back to it — a wake-up, a land or gate watch, or an approval card — and asks first so none of it is cancelled by accident. The refusal message names what is waiting. Choosing Archive anyway always goes ahead. The session's own scratch folder is never part of this check: it is temporary by design, so archiving deletes it either way. Off by default; turn it on in Settings → Sessions → Behavior if you want the check. An install that already had it on is switched off once by a one-time settings migration, and turning it back on afterwards keeps it on.

Show “not safe to archive” warnings on sessions

Draws a yellow shield on a session in the sidebar when archiving it would cut off work still coming back to it: a running background job, an armed wake-up, a land or gate watch, an unanswered approval card, an agent that said it is waiting, a live crew seat, or a queued message. Hover the shield, or the session name, to see which. Off by default, and while it is off Omniscio does not run the check at all. Turning it on brings the shields back within a few seconds; turning it off clears every shield already on screen. Separate from Warn before archiving loses work, which controls the prompt shown when you archive. Found under Settings → Sessions → Behavior.

Article Reader model

Which AI model is used to summarize articles you save in the Article Reader. Automatic (the default) inherits your global Utility AI model choice. Enter a specific model ID to override it for article summaries only. Found under Settings → Sessions → Per-feature AI models. Leave it on Automatic unless you want article summaries to use a different model than the rest of your utility tasks.

Aside cost cap (USD)

The maximum amount in USD that a single "aside" (a side-question spawned via Ctrl+B) is allowed to spend. If an aside is estimated to exceed this cap before it starts, it is aborted instead of launched. Range: 0–50; the default is $0.50. Set it to 0 to remove the cap altogether. Useful for keeping quick side-questions cheap even when your main sessions are running expensive models.

Ask Omniscio (in-app help agent)

When on, an "Ask Omniscio" virtual project appears in your sidebar — a friendly help agent pre-loaded with the Omniscio documentation and live app state. Ask it how to configure anything, what a setting does, or how to use a feature. On by default; turn it off to remove the tutor from your sidebar.

Ask Omniscio model

Which AI model Ask Omniscio answers with. Automatic (the default) uses a top-tier, Opus-class model — the reasoning is that a help agent which answers badly at the exact moment you need help is worse than no help agent at all. The list carries the Claude models plus, when you have another engine switched on, that engine's own models (DeepSeek, Kimi, and so on): picking one of those runs Ask Omniscio on that engine, which must be set up on this computer. If a model you picked later stops being available — the engine is turned off, or the model is retired — Ask Omniscio answers on Automatic instead and tells you so in the chat. Ask Omniscio runs on your own AI account; only if you have no usable account of your own does it fall back to a company-provided model, which is a cheaper model with a small per-conversation spend cap, and that fallback runs its own model whatever you pick here. Found under Settings → Sessions (advanced).

Plain-language replies

On by default. Adds a short instruction to every regular session telling the agent to write for a smart non-expert — plain everyday words instead of jargon or unexplained acronyms, start brief and add depth only if you ask, and clean scannable markdown that leads with the answer, uses clear headers and keeps bullets to one short idea each instead of walls of text. It is what keeps a fresh install from sounding verbose and developer-toned out of the box. Turn it off if you would rather agents write in their own default, more technical voice; nothing else about how they work changes. Settings → Features.

Auth pathway

Controls how native Claude sessions are credentialed. Four strictly isolated modes: managed ("Managed accounts") uses the pool of signed-in Anthropic OAuth accounts and saved API keys, with automatic load-balancing and failover — it is owner-only and is not offered on a customer install; passthrough ("Claude Code login") uses the credential from your own claude login terminal session stored in ~/.claude (Omniscio stores nothing); ownLogin ("Sign in to Claude") signs in exactly one Claude account inside the app, capped at one account and never pooled; apiKeyOnly ("Your API key") routes all sessions through your own Anthropic API key in a private slot. A customer is offered passthrough, ownLogin and apiKeyOnly, and an install that has not chosen yet falls back to passthrough — the credential nearly everyone already has. Set via the pathway picker during onboarding or in Settings → Accounts → Claude. Switching modes takes effect on the next session you start.

Auto-switch account on rate limit

When on, Omniscio automatically switches to another configured account if the current one hits a rate limit, rather than waiting for the limit to reset. This keeps sessions running smoothly when you have multiple Anthropic accounts in the managed pool. Off means sessions wait out the rate limit on their current account. Off by default since 2026-08-29 (fresh installs run a single account; earlier installs keep the setting they had). Only applies in managed auth pathway mode.

Auto-capitalize Team Chat

When on (the default), the Team Chat message composer automatically capitalizes as you type — the first word, the start of each sentence (after . ! ?), and the beginning of each new paragraph — while skipping URLs, file paths, and abbreviations. This is a free, local transform; it never calls the AI and never costs anything. Turn it off if you prefer to control your own capitalization in chat.

Apply updates automatically when idle

When on (the default) and an update has downloaded and is waiting to be installed, Omniscio applies it and restarts automatically — but only when no sessions are currently running and your computer has been genuinely idle. On by default since 2026-10-06, and a one-time migration switched existing installs on too: the reason is measured — an install running 24 h/day for 25 straight days never took an update it had already downloaded, because a waiting update is one a person has to remember. Turning this off is the way back to manual control: Omniscio then shows a nudge and waits for you to choose when to install.

Auto-lander repos

Per-repo configuration list for the in-source auto-lander feature. Which repos get landed comes from your Dev Pipeline repo selection — this list adds per-repo options on top of it: pausing one repo, setting the integration (target) branch it lands into, or turning on a dry-run mode that logs what would land without actually doing it. A paused or dry-run repo is only observed: it never gets a paid rescue session, its branches are never handed back to their authors, and its waiting queue never sets off the not-landing fixer. Adding an entry never removes a repo from the set, and an entry here does not add a repo your Dev Pipeline has turned off. To stop landing in one repo, turn it off in the Dev Pipeline panel; to stop landing everywhere, empty this list. The auto-lander master switch (autoLanderEnabled) must also be on.

Each repository also carries a Refresh generated indexes after each land switch (settleGeneratedIndexes), off unless set. Turn it on and, after a land, the auto-lander re-runs that repository's own generator scripts and commits the refreshed index documents — its contract, memory and postmortem indexes — straight onto the repository's trunk. That commit is the one time the lander writes to your trunk outside a land itself, so turn it on only for a repository that genuinely carries those scripts. It is not a switch you have to remember to set: at launch the app fills it in for you, once, for any configured repository whose folder really does carry Omniscio's generator scripts, and only when you have never set the key — a value you chose is never overwritten, and it is never enabled for a repository that lacks the scripts.

Approved developers

The GitHub logins whose pull requests may land on your computer without you looking first. A request from one of them lands as it always did. A request from anyone else is still fetched, checked and reported in full — it simply stops one step short of your main branch and waits for you, and the refusal it carries says what to do about it: add the login here, and it takes effect on the next minute with no new push from the author. A "tests passed" card on the pull request does not change this: the card says the code was tested, not that you trust who wrote it, so a request from anyone not on this list waits for you whether or not it carries one.

Matching is case-insensitive and trims spaces, because GitHub logins do. An author who cannot be read at all (a deleted account, an unreadable field) is treated as not approved — a merge gate that cannot identify who wrote something must default to not landing it, and the cost of that is one request waiting for you.

Each row also carries its own provenance: who added the login and the words behind it. A login you add here records owner (Settings); one a session asked for and you approved records owner (approval card); the ones that shipped with the app record who put them there and why, so the list can always be read back to a decision. That record is written by the app, never by whoever asked, and it cannot be rewritten afterwards — so the list never carries words you did not say.

Two things are deliberately outside this list, because neither has an author to check: branches you mark ready yourself, which the auto-lander lands, and a hand catch-up that merges your main branch.

Editing it: Settings → Auto-Lander (or the Dev Pipeline panel's All-settings screen), or by approving a card. An agent or a script may now ask for a login to be added or removed, and that request arrives as an ordinary approval card in your inbox: the card names the logins being added and the ones being removed, and names the session that asked. Nothing is written until you click — and nothing can clear that card on its own, not the "always allow" control (which is not offered here) and not the approval toggle in CLI Control. A whole-file settings backup still will not carry this list in either direction; the card and your Settings screen are the two roads.

Auto Model Router

Master switch for the Auto Model Router. When on, each new session picks its AI model automatically based on the task — using your routing rules, task complexity, and keyword signals — instead of always defaulting to your global default model. Off by default: nothing changes until you turn it on. Manage your routing rules in Settings → Model Routing. The router never moves a session whose engine was already named — an agent that asked for Claude keeps Claude on that session's own default model, so a worker pinned to an engine cannot end up on another vendor or a cheaper model. A "Not sure how to set this up?" card on the Model Routing settings page can launch a guided session that helps you configure routing rules based on your usage patterns and cost preferences.

Model routing rules

Your personal list of routing rules for the Auto Model Router. Each rule maps a task description (plain English "when to use" text, plus optional fast-match keywords) to a specific provider and model. New sessions matching a rule start with that model automatically. Rules are evaluated in order; the first match wins. Only active when autoModelRouterEnabled is on. Add, edit, and reorder rules in Settings → Model Routing.

Preview AI auto-replies before sending

When on (the default), an AI-written auto-reply generated by a channel automation is held as an approve/reject card in your inbox instead of being sent immediately. You see the exact reply text and where it will go before it reaches the recipient — protecting against a malicious incoming message tricking the AI into sending something harmful. Approve to send it or reject to drop it. Turn it off to send AI-written replies immediately without review. Plain template replies always send immediately regardless of this setting.

Auto-update tools

When on, Omniscio silently updates the bundled tools it can reach through npm — Playwright, AgentMail, and Repomix — in the background whenever new versions are available, without interrupting you. Claude Code and Codex are never updated silently: they ship as their own binaries and get an update-alert card instead (see Update alerts for Claude Code and Codex above). Git and GitHub CLI receive a notification rather than a silent update (they require more care). Off by default; turn it on to let Omniscio update Playwright, AgentMail and Repomix for you. Either way you are notified when updates are available.

Auto-wait period

How many minutes Omniscio lets a waiting session run silently before surfacing it to your inbox as needing attention. This single knob governs the prose-wait cap, the background-task hold, the "Keep waiting" countdown banner, and every auto-wait notice in the UI. Range: 5–240, default 30. Read internally as autoWaitMinutes ?? 30 when not set.

Enable Automation Builder

Turns the Automation Builder on or off. The Automation Builder is an in-app assistant that interviews you in chat, then builds, documents and tests an automation for you and sets it up to run on a schedule. Each automation it builds waits in your inbox for your approval before it can run. When this is on, the Automation Builder appears in the Automation group in the sidebar and as a tab in Quick Launch. Turn it off to hide those and keep its sessions out of your inbox. You find it in Settings > Features.

Default: on.

Channel filters (automations)

Master toggle for the Channel Filters / Automations feature. When on, your configured rules and AI processing run automatically against incoming messages, emails, and other channel events. Turn it off to pause all automated processing without deleting your rules; they will resume when you turn it back on.


Sign them back in automatically

Checks every few hours and drives the sign-in for any account that has dropped out, the same sign-in the button above runs. It exists only on an install that was already running a pool of several Claude accounts before the pooled pathway closed to new installs — a new install has no such panel and no such toggle, and signs back in by hand from the Claude tab in Settings. On an install that has it, the setting needs the Omniscio Chrome extension (My Real Chrome Lite) turned on in Settings, Features and connected in Chrome. You still do the one click Cloudflare insists on. Off by default, and nothing runs at all while it is off.

While no extension is connected every check skips quietly, and the panel above shows a Set up Chrome button instead of a sign-in it cannot run. See Sign back in to Claude when you are logged out.


Let agents offer a key for you to review

Lets an AI session offer a key it already holds so you can review it, edit it, and press Save to store it. Nothing is saved unless you press Save. On by default; turning it off still lets an agent pop the prompt, it just cannot pre-fill a value.


Let agents save API keys you give them

When you hand an AI session on this computer a key, password or web address, it saves it straight away — no box, no Save click — through the same encrypted storage your own Save uses. AI provider keys are checked with the provider first, and a rejected key is never saved; if the provider can't be reached the key is saved and the session is told it wasn't checked. Ordinary non-secret settings still ask you, and plug-in tool servers, cloud sessions and outside API keys always get the review box instead. Off by default, and only you can turn it on — no agent or command-line token can switch it on for itself. Needs "Let agents offer a key for you to review" on as well.


Which key pays for each AI helper

The order Omniscio tries your provider keys in to pay for the built-in AI helpers — session titles, reply suggestions, reply drafts and summaries — top to bottom, first row that resolves wins. There are four rows: openrouter, deepseek, anthropic and allowance. The default is exactly that order, so your own keys are tried before the Omniscio allowance. A partial or missing list can never strand a helper: whatever is omitted is appended in the default order. Reorder the rows in Settings → API Keys. List of rows; accepted values: any order of openrouter | deepseek | anthropic | allowance, each at most once.


Pin one key to one AI job

Sends one AI-helper job to a specific key row regardless of the order in "Which key pays for each AI helper". The four jobs are titles, suggestions, replyDrafts and summaries; each may name any of the four rows (openrouter | deepseek | anthropic | allowance). Empty by default, meaning every job follows the order above. Set one in Settings → API Keys when a single job should always bill a particular key.


Unanswered approvals

Only shown while the Safety checks Lab switch is on (Settings, Lab). Decides what happens to an approval card you have not answered. The default, Wait for me (never time out), leaves the card up until you answer. Choose Refuse after 5, 15 or 30 minutes and an unanswered request is refused when the time runs out, the session is told, and a later answer is ignored. Sessions nobody is watching (scheduled jobs, recipes, automations, hidden sessions) are not affected by this choice: their asks are refused at once. Stored as off, 5, 15 or 30. Find it in Settings, Safety checks. See safety-checks.md.


Anthropic admin key (vendor key watch)

An optional Anthropic organisation admin key the vendor key watch uses to read Anthropic's usage report for the API key Omniscio holds. Anthropic only reports spend to an admin key, and an admin key can manage your whole Anthropic organisation — so it is stored encrypted, kept in the main process, used for reads only, and never shown in a card, a log, the Stats rows or the command-line route. Leave it empty and Anthropic shows as "not watched — needs an admin key". Create one at console.anthropic.com → Settings → Admin keys. Secret text; accepted values: string (≤500).