Agent Permission Level (choose how much agents can do before Omniscio asks you)
One setting that decides how much freedom your agents get before Omniscio stops and asks you: Read-only, Guarded, Autonomous, or Full trust. You choose it during onboarding or any time in Settings, and it applies to your Claude Code sessions and your Codex sessions right away, including Codex sessions already running.
On this page
What it is
Omniscio runs agents in the background, unattended. Left to itself, an agent would edit files and run commands without ever asking — which is fast but gives it a lot of freedom. Agent Permission Level is a single setting that lets you decide how much freedom your Claude, external (Kimi Code / Hermes / DSH) and Codex sessions get before Omniscio stops and asks for your approval (Codex maps each level to its closest built-in policy — see Which agents it controls). It has four levels, from most cautious to most trusting, plus an optional fifth, Smart, that appears only when the Safety checks Lab switch is on (see Safety checks):
- Read-only — the agent can read your files and propose changes, but it asks you before writing any file or running any command. Maximum control.
- Guarded (the default for new installs) — the agent automatically makes file edits inside your project, but asks you before running a shell command, editing its own config files (
.claude/,.git/,CLAUDE.md, and similar), writing a file outside the project folder, or using a connector (MCP) tool or any other tool Omniscio cannot classify — those can send, post or delete on an outside service. - Smart (only with the Safety checks Lab switch on) — sits between Guarded and Autonomous. It decides exactly as Guarded does, but each time Guarded would ask you, a small background model judges the request instead: clearly safe is allowed, clearly dangerous is refused, and anything unsure (or any failure, missing key, timeout or spent daily limit) comes back to you. It can never allow a command on the dangerous list. Each verdict leaves a one-line reason in the session. With the Lab switch off, a Smart session behaves as Guarded.
- Autonomous — the agent automatically makes edits and runs commands inside the project; it only asks before writing files outside the project.
- Full trust — the agent does everything on its own, with no permission prompts — with two exceptions: the safety floor and the short always-ask list, both below.
Below Full trust (Read-only, Guarded, Autonomous), Omniscio always asks before an agent touches a genuinely sensitive file (SSH keys, cloud credentials, system files like /etc/passwd, shell startup files, and Omniscio's own ~/.amc machine-state folder). Full trust is the deliberate exception — it auto-approves everything else without a prompt. One thing Full trust does not wave through is the always-ask list: writing to your SSH keys (~/.ssh/), your cloud credentials (~/.aws/, ~/.config/gcloud/, ~/.azure/, ~/.kube/config, ~/.docker/config.json), a shell startup file (~/.bashrc, ~/.bash_profile, ~/.profile, ~/.zshrc, ~/.zprofile, a PowerShell profile), or a git hook (.git/hooks/ in any repo) still asks at every level, Full trust included. Only writes ask — reading those files or listing the folder stays silent. An external agent (Kimi Code, Hermes, DSH) is the one case where a sensitive file is refused at every level, Full trust included — see Limits and edge cases. One rule never relaxes at any level: sessions started by an inbound email are locked down harder (they can't perform risky operations at all), and — as a second guard for that same untrusted-inbound case — an email-triggered session keeps the sensitive-file prompt even under Full trust.
The safety floor above Full trust
A small set of protections sits above every level, Full trust and Smart included, and cannot be approved away:
- Catastrophic commands (wiping a disk, a home folder or a whole system, fork bombs) are always refused. This is on by default and has its own switch in Agent permissions.
- Never-allow rules you write yourself are always refused (Safety checks Lab switch on).
- Writes into credential folders (SSH keys, cloud logins and the like) are refused (Safety checks Lab switch on).
These are described in full, with their defaults and limits, in Safety checks. They cover Claude sessions and the external agents; Codex receives only the command rules, and Pi is not reached.
Where to find it
Where it lives in the UI
You choose your level in two places:
- During onboarding — a step called "Agent Access" (in the personalize part of setup, right before the theme picker) shows the four levels as cards. It starts on Guarded; tap a different card to change it.
- Any time afterward — Settings → Permissions, a dropdown labelled "Agent Permission Level", sitting just above the "Auto-Approve Plans" and "Auto-Approve Config File Writes" toggles. Each option spells out what it does; pick one and it applies immediately.
The same Agent permissions area also holds these rows:
- Refuse catastrophic commands, Refuse asks nobody is there to answer and Remove secrets before text leaves this computer — three always-on switches (default on).
- Never-allow rules — your own list of commands and
path:patterns that are always refused (needs the Safety checks Lab switch). - Suggested rules — rules offered from commands you keep allowing by hand; Add as a rule applies one, nothing is added on its own (needs the Lab switch and its suggest-rules switch).
When a level makes Omniscio pause for your approval, the session turns amber ("Needs You") and a small card appears with Allow / Deny buttons. The card now tells you why it's asking and what it's asking about — for a command it shows the actual command line (e.g. "Claude wants to run a command: npm run build"), and approving it runs that exact command.
Allow (and its Y / Enter shortcuts) answers only this request. A small arrow attached to the right of Allow opens the always-allow panel, headed Allow and stop asking about, with three short choices: what (a precise rule for the command's leading words such as "git" commands — or, for a non-command ask, Every <tool>), where (This session only, This overseer — shown only when the asking session is an overseer, and covering whichever session holds that overseer post now or later — Projects with the current project ticked, or Everywhere) and for how long (1 hour, Today — which ends at midnight — 1 week or Forever — the default, so a standing permission is the rule and one-off gating the exception). A one-line summary says exactly what will happen, and Allow & stop asking (or Enter) applies it and approves this request, with an Undo toast that also puts back any rule it replaced. With the Safety checks Lab switch on, the arrow also offers Allow for this session (shortcut Shift+S), which stops repeat asks for that command until the session ends. The broad Agent Access change is a small link at the foot of the panel. Broad choices (a whole tool everywhere, or the level change) confirm first, an open panel closes when the next request arrives, and the arrow is hidden entirely on asks no setting can stop (sensitive files, a Clean Room escape). A timed rule switches itself off at its deadline — on cloud machines too — and Settings → Permissions → Always-allow rules lists every rule with where it applies and how long it lasts.
Once you Allow or Deny and the session has nothing else pending, Omniscio moves you straight to the next session that needs you — the same way sending a reply does, following your Post-send navigation setting (Settings → Sessions → "After I send a message…"). If the agent has stacked several commands on one session, you're shown the next one first and only move on once they're all handled — so triaging approvals flows exactly like triaging replies.
How it behaves
New installs vs. existing installs
The default is deliberately different depending on whether Omniscio is new to you:
- A brand-new install starts on Guarded, so a first-time user isn't handing an agent full freedom before they understand it.
- An install that already exists (you've onboarded before) is pinned to Full trust the first time it starts up after this feature ships — so your running agents keep behaving exactly as they did, with no sudden wave of approval prompts. You can switch yourself to a stricter level whenever you like in Settings; nothing forces you to.
This "keep existing users where they were" migration runs exactly once and never overrides a level you later choose yourself.
Which agents it controls
This setting governs Claude Code sessions (and the Claude-compatible engines like DeepSeek and GLM that run through the same path), the external coding agents that use the shared approval channel (Kimi Code, Hermes and DSH — their tool calls now ask you at exactly the level you chose, instead of being waved through), and, since 2026-09-23, new Codex sessions. Cursor and Gemini keep their own permission systems: Cursor decides permissions server-side in your Cursor account, and Gemini manages its own approvals.
Codex has its own built-in permission system, so Omniscio picks the Codex policy closest to your level:
| Your level | What a Codex session does |
|---|---|
| Read-only | Cannot edit files, and asks before every command |
| Guarded | Full machine access, but asks before every command or file edit |
| Autonomous | Works freely inside the project folder, internet included (npm install, git push); asks only before writing outside it |
| Full trust | Full machine access, never asks |
- Codex can have its own level. Settings → Accounts → Codex → Codex permission level defaults to Follow Agent Permission Level. Pick another option there to run Codex at a different level from your Claude sessions.
- Upgrading switches the old default once. An install still on Codex's old default (Ask before every command) moves to Follow Agent Permission Level the first time the new version starts. A different Codex level you picked on purpose is kept.
- Changes reach running Codex sessions. A running session picks up your change on its next message. In the middle of a task, Full trust approves the next request by itself, and Read-only refuses the next file edit. A task already running at Full trust keeps going until it ends, because Codex stops sending requests to Omniscio at that level.
- A card already on screen is answered for you when your new level would not have asked. Switch a session to Full trust while an approval card is waiting and the card clears, the request is approved and the agent carries on — the same as if you had approved it, which is what you just told it to do. Change to any other level and the card stays exactly as it is: answer it once, and that session follows the new level from then on. Staying is the safe half, so a card is never declined for you and a card that still needs your judgement never disappears.
- The card's Agent Access row shows the level that request was really judged at. On an ordinary session it is the dropdown, and changing it changes your setting. On a session held at a stricter level than your setting (a session started with its own lower level, a concierge chat, a guest's turn) it reads, for example, Agent Access: Guarded (held for this session) with no dropdown, because changing your setting would not loosen that session. On a session running on another of your computers it reads Agent Access on <that computer>: Guarded — that computer's own level, which it judged the request by — and no "stop asking" level change or always-allow rule, since those would only change this computer. When that computer runs a current Omniscio the row is a dropdown again, and picking a level sets that session's level over there: Full trust there means no card for anything Full trust allows, and the card you are looking at clears if the new level allows it. That computer decides: it never goes above its own Agent Access, it refuses to loosen a session if it is set to Only me, at this computer, it only loosens a session that was started from another of your computers (never one an email, a bug report, a task, an inbound message, a Clean Room or the concierge started), a card that only a person at that computer may approve is never approved by the change, and tightening is always allowed — it covers what the session asks about at once and everything else once the session restarts. Only you can do this from the app or your paired phone — agents cannot. A computer too old to say which level it used shows no level at all rather than a guess.
- A session locked to a stricter level stays locked. Your level and the Codex level can only make that session stricter, never looser.
- Claude-only protections stay Claude-only. The sensitive-file check and the approval card's "always allow" options don't apply to Codex, because Codex answers its own approval requests.
Limits and edge cases
- Guarded is deliberately chatty for command-heavy work. Because it asks before every shell command, an agent that runs a lot of commands will generate a lot of approval prompts. That's the point of the tier — control — and it's why command-light first-time use fits Guarded while power users tend to pick Autonomous or Full trust.
- The two "Auto-Approve" toggles still apply on top. "Auto-Approve Config File Writes" can still force config prompts at the more trusting levels; "Auto-Approve Plans" is a separate thing (it governs plan-mode approval, not tool permissions) and this setting doesn't change it.
- Reads are always fine at Guarded — including reading a config file. Guarded only asks before changing config, not before looking at it.
- If Omniscio can't tell where the project folder is (a rare session with no resolved working directory), an out-of-project write can't be judged, so it's allowed rather than blocked — Omniscio never blocks work just because it's unsure. Below Full trust, the sensitive-file protection still applies.
- A session running on another machine (over SSH) is judged by its folder on THAT machine. "The project" is the folder Claude reports when it starts there, plus that machine's
/tmp— never the project folder on your computer, whose paths mean nothing on the remote. Until Omniscio has heard that folder, every write from the remote session asks: unlike a local session, a remote one is never let through on a guess. - The sensitive-file check reads the whole command, so it can misfire below Full trust. It scans a shell command word by word, which means a command that merely mentions a sensitive path — a
git commitmessage that talks about~/.bashrc, for example — can raise the prompt even though nothing touches that file. On a Claude session at Full trust this never happens, because that broad check doesn't run there — only the always-ask list does, and it fires on a write shape (a redirect,cp,mv,rm,tee,sed -i) rather than on a mention. On an external agent (Kimi Code, Hermes, DSH) the broad check always runs and a sensitive file is refused outright — see the next bullet. (It reads the paths a tool call names, not the text it is about to write, so a note that merely mentions a sensitive filename is not refused.) - An external agent's approval is the same setting, with two differences you should know. Kimi Code, Hermes and DSH ask you at your level like Claude does, but a sensitive file (SSH keys, credentials, system files) is refused outright at every level — Full trust included — rather than being offered as a card you could click through; an external agent was never granted that bypass. And their cards offer the Agent Access level change but not the fine-grained "commands starting with
git" rows, because those rules are read only by the Claude gatekeeper and a rule written there would quietly do nothing. - If Omniscio genuinely can't tell what an external agent is asking to run, it asks you rather than assuming it is harmless — at Guarded and Autonomous alike. Only Full trust lets an unidentifiable tool call through, which is what makes the card's "Agent Access → Full trust" option an honest one.
- A session can be pinned stricter than your global level. The Chief of Staff (Simple Mode) session runs at Read-only — it asks before anything that is not a lookup — and the guided app tour runs at Guarded, on purpose, so both keep the sensitive-file prompt even when you're on Full trust. A pin can only ever make a session more cautious, never less. A workflow that an incoming message or email started pins its AI steps the same way — see Workflows, "What an AI step is allowed to do".
- A pinned session asks even when your computer would let it run. Claude Code normally answers from your own Claude settings first (the allow list in
~/.claude/settings.json, a "bypass" default mode) and only asks Omniscio about what those leave open — so a pin used to be skipped for anything your computer already allowed, including on a linked computer. A pinned Claude session now launches with its own ask rules, so every command, edit, web fetch, MCP tool and file read reaches Omniscio, which still lets reads and your safe commands through on its own. If those rules cannot be written, the pinned session does not start rather than run unpinned. Not yet covered: asides and SSH-machine sessions.
One session at a time
The shield is hidden by default. Settings → Sessions → Composer buttons carries a Permissions (shield) switch, off unless you turn it on, so the message box keeps that space free. While it is off there is no way to change one session's level on its own: Settings changes the level for every session.
The shield button in a session's message box (desktop app) sets the level for that session only. It can make the session stricter or looser than the global level, up to Autonomous; Full trust is chosen only in Settings, and the button offers it only when the global level already is Full trust. Every floor still applies (an email-started session, the concierge, a guest's turn, Clean Room, sensitive files). Cloud, remote and shared sessions can be made stricter but not looser. Each change adds a line to the transcript. The phone shows nothing to change, and it can no longer change the global level, the allow rules, the Codex approval mode, plan auto-approval or config-file auto-approval either.
Related
Related pages
- settings-patch-approval.md — how an external caller (CLI/agent) proposing a settings change is queued for your approval; distinct from this in-session tool-permission gate.
- share-cli.md — the local CLI control server, where this setting is also readable/writable via
/settings. - agent-firewall.md — the other control on what reaches an agent: a screening layer over what a tool hands back, with its own warn, alert and block levels.
Last verified 2026-11-11