Support

Approvals & Safety Guide

How CellCog decides what an agent may do on your behalf: the three threat levels, your auto-approve thresholds per surface, pending approvals in chat, and standing approvals for AI employees.


The Model in One Paragraph

Every action an agent takes on your computer (Cowork), in a browser (Browse my Chrome or a cloud browser), or through a connected app (Tools) is classified by the agent as safe, moderate, or dangerous, with a plain-language summary and, for elevated levels, a one-sentence reason. You set an auto-approve threshold per surface. Actions at or below your threshold run immediately; actions above it appear as a pending approval in chat and wait for you. AI employees work on the same model with two additions: their level is at least Moderate, and you can grant standing approvals for classes of dangerous work they need repeatedly.


The Three Threat Levels

Threat tracks data, not the kind of command. The same click or command can be safe, moderate, or dangerous depending on what it touches. The agent answers two questions in order; the first yes wins:

  1. Does this change data, act on your behalf externally, or do something you cannot easily undo? Then it is dangerous.
  2. Does this read your private data? Then it is moderate.
  3. Neither? Then it is safe.

When in doubt the agent is expected to pick the higher level.

table
LevelCowork (your computer)BrowsersTools (connected apps)
SafeListing files, running tests or builds, reading and editing files inside a version-controlled working directoryOpening a public page, searching a public site, expanding a menuListing files, reading public profiles, reading your own calendar or project data
ModerateReading files outside the working directory, environment variables, shell history; editing files in a project with no version control and no backupSnapshotting a logged-in page (inbox, private repo, banking), reading cookies, listing or adopting your personal tabsReading highly private data: full email content, payment history, private messages
DangerousInstalling packages, deleting files, editing shell profiles or system files, pushing to a remote, anything touching productionSubmitting forms, sending email, paying, deleting, writing cookies, downloading filesAny write or externally visible action: sending email, posting, creating invoices, changing CRM records

Two special cases:

  • Recovery matters on your computer. Inside a git repository, edits are recoverable and therefore safe. Without version control or a documented backup, the agent treats edits as moderate and may ask how you back the project up.
  • Your own MCP servers. Tool calls to servers you registered yourself are always at least moderate, because a third-party server's descriptions are untrusted. The approval card marks these tools as served by your own MCP server.

Auto-Approve Thresholds

You choose how much runs without a click, separately for each surface:

table
ThresholdWhat runs automatically
NoneNothing. Every action requires your approval.
SafeSafe actions only. Private reads and anything that changes data still ask. This is the default for every surface.
ModerateAlso actions that read your private data. Anything that changes data or acts on your behalf still asks.
DangerousEverything the agent declares, including actions that change data or act on your behalf. Use only when you fully trust the task.

Where the account default lives

table
SurfacePageSetting
Coworkcellcog.ai/coworkAuto-approve threshold
Browse my Chrome (and cloud browsers)cellcog.ai/cowork/browseAuto-approve threshold
Toolscellcog.ai/connectors/toolsAuto-approve tool calls up to

Your profile page shows all three side by side under Auto-approve up to (Terminal on my computer, Browser control, Connected app actions), with a link to change each.

Per-chat overrides

Each surface chip in the chat editor (Cowork, Browse, Tools) carries a colored dot showing the effective level for that chat. Hover the chip and click the dot to set a level for this chat only; Reset returns the chat to your account default. New chats always start at the account default. You can also change the account default from the same panel.


Pending Approvals

When an agent sends an action above your threshold, an approval panel opens on the right side of the chat. Each command card shows:

  • The command type (terminal command, file read, file write or edit, browser action, or tool call)
  • A colored threat badge (safe, moderate, dangerous) and the agent's reason when elevated
  • The agent's summary in plain language
  • The details: the exact command, the script, the file change, or the tool's arguments

Your choices:

  • Approve or Deny each command. A denial can carry a reason, which the agent reads and uses to adjust.
  • Approve All or Deny All for the whole batch (Deny All asks you to confirm)

A batch runs in the order the agent sent it, and starts only after every command in it has been decided. Commands at or below your threshold in the same batch are marked Auto-approved and run without waiting. In a regular chat the agent waits for your decision as long as it takes.

If an agent forgets to classify an action, or declares an elevated level without a reason, the action does not run at all. The card shows Will retry and the agent re-issues it with proper metadata on its next turn; nothing is needed from you.


AI Employees

AI employees work autonomously, often while you are away, so their approval model has three additions.

Their auto-approve level starts at Moderate

An employee's level per surface is Moderate (the default) or Dangerous. None and Safe are not available, because an employee cannot stop and wait for every read. You set the level when you hire the employee and can change it later from the chips in the employee's chat.

What happens above the level

When an employee wants to do something above its level, the platform checks, in order:

  1. A standing approval. If the employee quotes an active approval you granted for that class of action, it runs immediately.
  2. You are present. If you are in the employee's chat, the action appears as a pending approval like any other. If you do not decide within 15 minutes, the action fails and the employee's shift continues with its other work.
  3. You are absent. The action fails immediately with guidance, and the employee asks you for an approval through a task.

Standing approvals

A standing approval is your yes to a class of dangerous work, described in one sentence, for example "Send up to 5 cold outreach emails per day" or "Post the weekly update to the team channel". How it works:

  1. The employee asks you, in chat or through a task assigned to you, spelling out the exact class, limit, and duration
  2. You confirm in chat
  3. The employee records the approval, which gets an id like AP-7Q2K4M. Approvals have no expiry by default; a temporary grant can carry a number of days.
  4. From then on, actions in that class run without asking, and you see the id quoted on them

You can see every active approval in the Approvals app on the employee's computer, with its description, grant date, and expiry. That view is read-only: approvals are granted and revoked in conversation. To revoke one, tell the employee and it removes the approval; expired approvals stop working on their own.

The platform checks that a quoted approval exists, is active, and has not expired. The wording of the approval is the contract the employee is expected to honor; if it oversteps, tell it and revoke.


Hiring is never covered by an approval. An AI employee (or a chat) can only propose a hire: the New AI Employee form appears filled in for you, and your Hire click is the hire. No standing approval, id, or setting lets an agent create an employee on its own.

What Never Runs Without You

  • Dangerous actions when your threshold is below Dangerous, unless an employee holds a standing approval for that class
  • Actions with missing or malformed classification (they are rejected before anything executes)
  • Reads of credential locations on your computer, such as ~/.ssh, cloud CLI configs, .env files in your home directory, and shell profiles. These are blocked regardless of threshold.
  • Anything on your computer or in your Chrome when the desktop app is not running

The Yellow Browser Banner

When an agent drives your Chrome through Browse my Chrome, Chrome shows a yellow banner on the controlled tabs: "CellCog Browse started debugging this browser." with a Cancel button. This is Chrome's own safety indicator; CellCog cannot hide it. It appears while the agent is executing a batch and clears shortly after. Click Cancel to revoke control of that tab immediately. Cloud browsers show no banner, because they are the employee's own Chrome on CellCog's servers.


What Data Leaves Your Machine

  • Cowork: files the agent reads and command output it receives are sent to CellCog and become part of the chat's context. Secrets such as API keys, tokens, and private keys are redacted from output before it leaves your machine. Credential paths are blocked entirely. A local history of executed commands is kept on your machine and cleared after 90 days.
  • Browse my Chrome: page content, form values, and screenshots of tabs inside the CellCog tab group. Password and credit-card fields are redacted before anything leaves Chrome. Tabs outside the group, your history, bookmarks, and saved passwords are never read.
  • Cloud browsers: nothing from your machine. The browser and its logins live on CellCog's servers.
  • Tools: the results a connected app returns flow through CellCog to the agent.
  • All traffic between your machine and CellCog is encrypted in transit.

Frequently Asked Questions

Why does the agent choose the threat level instead of CellCog?

Only the agent knows what a command is about to touch. Its declaration is what your threshold gates, and the card shows you both the level and the reason so you can judge it. Underclassification shows up on the card; when in doubt the agent is expected to escalate.

What is the safest setup?

Threshold None on every surface: every action waits for your click. Most people settle on Safe or Moderate and keep Dangerous for chats they fully trust.

Does a per-chat override change my account default?

No. Overrides apply only to that chat and are never copied to new chats. Changing the default from a chip does update the account-wide default.

Can an agent talk its way past my threshold?

No. The threshold is enforced by CellCog, not by the agent: an action above it stops and waits for you no matter how the agent describes it. The card shows you the level the agent declared and its reason, so an underclassified action is visible when you review it.

What happens when an employee's action times out?

After 15 minutes without a decision, the action fails, the employee is told to get a standing approval or retry while you are present, and its shift continues with other work.

How do I revoke a standing approval?

Tell the employee in chat. It revokes the approval, and the Approvals app on its computer stops listing it. Timed approvals expire on their own.

Do approvals apply to the employee's email?

Employee email works differently: there is no approval button on outgoing mail. Your dangerous-operation policy for the role, set during onboarding, decides what the employee sends on its own and what it brings to you first. See the AI Employees Guide.


Related Guides