Permissions
Permissions decide what your agents may do on their own, and what has to stop and ask you first.
This is the most important setting in C5. It is worth ten minutes.
Human teammate access & user permissions (RBAC) Looking to control what human team members can see, edit, or execute in the C5 interface? This page covers autonomous agent permissions. To configure role-based access control (RBAC), feature locks, and server controls for human teammates, see Access & user permissions (RBAC) in Settings.
The ladder #
Think of it as a ladder, from safest to most powerful:
| Level | What it covers | Default |
|---|---|---|
| Read | Looking at files and pages. Changes nothing. | Allowed |
| Write | Creating and editing files. | Asks you |
| Run | Running commands on your computer. | Asks you |
| Send | Sending anything off your machine — email, posts, calls. | Asks you |
| Delete | Removing files or data permanently. | Asks you |
| Spend | Anything that costs money beyond normal model usage. | Asks you |
| Control | Driving your screen, mouse, and keyboard. | Off |
Reading is safe: worst case, an agent wastes time. Deleting and sending are not: worst case, something is gone or something private went somewhere public.
Note Writing asks you by default, wherever the file is. A write goes ahead without asking only where you have said so: a folder you listed under the write-allowed paths in Settings → Extensions (empty until you add to it), an Always Allow you gave that agent for that tool, or the tool set to On in Settings → Extensions. An orchestrated run you have given filesystem autonomy can also write inside its own task workspace without asking, and only there.
Settings → Locations is a different setting and does not grant anything. See the warning further down this page.
When an agent asks #
You get a clear message saying exactly what it wants to do — which file, which address, which command. Then you choose:
- Approve Once — just this time.
- Deny — not this time; the agent finds another way or stops.
- Always Allow — do not ask again for this agent and this tool. It covers only the agent that asked: every other agent is still asked, and an Always Deny you gave another agent stands. Where the prompt offers it, you can limit this to the one file or address it asked about. Under "Always" lasts you pick how long it holds: Forever, 1 hour or 10 uses.
- Always Deny — refuse this agent and this tool from now on, without asking (or only that file or address, if you chose that).
An "Always" is about this agent. It never sets the tool's switch under Master Tool Authorizations (below) to On, which would let every agent use the tool without asking. The one move it can make is the other way: an Always Deny for the whole tool turns a switch that is On to Off, because while it is On every agent runs the tool, this one included, whatever its own rule says. A switch at Prompt or Off is left as it is, and an "Always" for one file or address moves no switch.
Under Prompt, C5 still asks. A tool at Prompt asks before every call, whatever an agent's own rule says, so your answer tells you so: read_file is set to Prompt in Master Tool Authorizations, which asks before every call to it, so C5 will still ask. Your “Always Allow” for researcher is saved and applies if that switch is set to Off in Settings › Extensions.
If C5 could not save part of your choice, a message says which part, for example “Always Allow” for read_file was not saved (HTTP 403), so C5 will ask again next time. or, for an Always Deny that had to turn the switch off, The Master Tool Authorizations switch for read_file was not changed (…); it is still On, so every agent still runs read_file without asking. Set it to Off in Settings › Extensions. When C5 cannot read the switches, it changes none of them and says so, so the other tools' switches are never reset. Your answer to the call itself goes through either way.
Only an administrator at the machine can answer. Anyone else sees the request, with Only an administrator can answer tool requests. in place of the buttons (on a paired device, You can answer tool requests only on the machine running C5.), and a Close button, which Escape presses too. Close only takes the prompt off that screen: nothing is sent, and the request waits for an administrator until it times out. In the notification list's Tool Activity, a call waiting for an answer shows its Approve and Deny buttons to an administrator only; a teammate who may act on that list reads Only an administrator can approve or deny tool requests. instead.
Tip Be slow with Always Allow. It is the choice people regret. Approve Once costs you two seconds and keeps you in the loop.
When an agent lacks a tool #
An agent sometimes needs something C5 has not been set up with — a local tool, or a way to make video or audio. With Tool & Web Escalation on, it can ask for one. The checkbox is in Settings › Tools, in the Tool Use panel; it is off by default, takes effect when you press Save, and only an administrator at the machine can change it.
The request comes as a Tool Authorization Required dialog, marked Tool Capability Escalation, Free Web Resource Escalation or System Control & Web Escalation, and saying why: Agent … is requesting tool escalation because a local capability is unconfigured. It shows the Missing Capability, the Free Web Resource where there is one, and the Escalation Target. You answer it like any other request. What an approval grants is fixed:
- A local tool (
request_tool_escalation): that tool, for that agent, for up to 10 uses within 30 minutes. - A public web resource (
request_web_resource_escalation): reading that address, and — unless the agent asked without it — System Control for that agent: C5's tools that drive the screen, mouse and keyboard, each for up to 25 uses within 30 minutes. Approve this only if you would let the agent drive your screen.
An approval never overrides a rule you set:
- An Always Deny stays. A tool that agent was told Always Deny — for the whole tool, or for the address or command it asked about — is left denied; the agent is told so, and not to ask again. A web escalation still grants the System Control tools that are not denied.
- An approval of nothing but denied tools grants nothing. When
read_urland every System Control tool it would grant are at Always Deny for that agent, the agent is told the approval granted nothing, and System Control stays off for its task. When only the System Control tools are denied, it getsread_urland is told that System Control was not granted. - An Always Allow you gave stays as you gave it, with no budget put on it.
- A request about one address or command grants only that, where the tool works on one; otherwise the whole tool.
- Asking again does not top it up. While a grant from an earlier approval still has uses and time left, a new approval keeps it as it is; a new budget starts only once it is spent or expired.
When its uses or its time run out, the tool asks again. Spent and expired grants are cleared away every night.
With the checkbox off, an agent's request for escalation is answered that escalation is off, and you are not asked — unless the request tool is set to On or Prompt under Master Tool Authorizations (below), or you gave that agent an Always for it.
Setting the level #
Which tools ask you first is set per tool, in Settings → Extensions, and by your answers to the approval prompt above. Settings → Extensions gives every tool one of three master positions:
| Position | What it means |
|---|---|
| On | Every agent may use this tool without asking. |
| Prompt | Agents must ask you each time, even an agent you gave Always Allow for the whole tool. An Always answer for one file or address still applies to that file or address. In a task run nobody may be there to answer: once the Authorize-Use Timeout in Settings › Tools passes, the step is suspended, not refused, and resumes when you approve it. A chat nobody answers is denied at the timeout. |
| Off | No override: the tool's own per-agent rules decide, and anything you have not answered with an Always still asks you. It is not a block. To stop a tool, choose Always Deny when it asks. A tool nobody has set is at Off, and the panel shows it so. |
Master Tool Authorizations, in Settings → Extensions, is where those positions are set, one switch per tool, listed by the tool's own name. Besides the tools agents already have, it lists:
| Tool | Group | What it does |
|---|---|---|
render_video |
Media — generate | Makes a video with captions from slides, on this machine |
fetch_web_image |
Media — generate | Downloads one image from the web into the task |
create_file |
Files — write | Another name for write_file, with its rules |
request_tool_escalation |
Tool & Web Escalation | Lets an agent ask for a local tool it lacks |
request_web_resource_escalation |
Tool & Web Escalation | Lets an agent ask for a public web resource, and System Control with it |
On for request_web_resource_escalation approves every such request without asking,
System Control included. Only an administrator at the machine can change a switch (see
Settings only an administrator can change).
Settings → Autonomy (headed Autonomy Policies) sets how far agents may go without you for each type of work: under Work Type Thresholds, the minimum autonomy level that work needs, from Level 1 (Tools) to Level 5 (Deploy), and what happens when an agent's level is too low (Downgrade or Block Task). Agent Overrides forces that choice for one agent and type of work, and Escalation Notifications lists the tasks that were blocked or downgraded this way.
Between the two, you decide exactly how much rope your agents get.
Settings only an administrator can change #
Some settings each take a prompt away from every agent, or let C5 do more with nobody asked. Only an administrator can change them, and only at the machine running C5: a phone or tablet paired through Remote Control is refused, even when it is signed in as an administrator. Anyone else sees them as they are, held back, with the reason beside them (below).
| Setting | Where it is |
|---|---|
| Tool & Web Escalation | Settings › Tools, in the Tool Use panel |
| System Control (computer use) | Settings › Tools |
| Egress Data-Loss Prevention: its mode (Off, Warn, Redact, Block) and Scan PII | Settings › Tools |
| Strict Path Boundary and Enable Root Level Terminal | Settings › Tools |
| FFmpeg location | Settings › Tools (see Pictures and video) |
| Master Tool Authorizations | Settings › Extensions, on the Saved Tool Authorizations card |
| Permitted Directory Paths, read and write | Settings › Extensions, on the Saved Tool Authorizations card |
| State Machine HITL Decay Threshold: how many approvals of a recipe step skip its sign-off from then on | Settings › Autonomy, under Global Rules |
| Guardian Vetting: the guardian's check of agents' tool calls | Settings › Budget, under System-Level Prompting |
| Back up secret files | Settings › Integrations, on the GitHub card, under C5 Data (see Backing up to GitHub) |
C5 also strips credentials from the environment of every command an agent runs. That has no switch in Settings, and only an administrator could turn it off.
Guardian Vetting is off only while it is saved off. If C5 cannot read the setting, the guardian goes on vetting.
What a teammate sees. For anyone but an administrator at the machine, these controls are greyed out, with the reason beside them:
| Where | What it says |
|---|---|
| Settings › Tools | Every panel is greyed out and Save is held back, with Only an administrator can change tool settings. under the heading. The list of rate-limited agents at the foot of the tab, which only reads, is not |
| Settings › Extensions | The Saved Tool Authorizations card is read-only, with no Save for it and Only an administrator can change tool authorizations. under its heading |
| Settings › Autonomy | The decay slider, with Only an administrator can change the HITL decay threshold. |
| Settings › Budget | The Guardian Vetting switch, with Only an administrator can change Guardian vetting. |
| Settings › Integrations | The Back up secret files switch, with Only an administrator can change whether secret files are backed up. |
On a paired device each reads You can … only on the machine running C5. instead, for example You can change tool settings only on the machine running C5. On Settings › Budget and Settings › Autonomy only what you changed is sent, so these settings never block a teammate's other changes there.
What is recorded. Every change an administrator makes to one of them is recorded in the audit log, with who made it, and so is every attempt from a paired device. A save that changes nothing — the same value, or the same list in another order — is not recorded as a change. See Audit trail.
Two settings work one way round. A teammate can tighten Remote Control's restrictions but not loosen them (see Who can change these switches), and can turn Voice on remote devices off but not on (see Voice from a paired phone). Once the setting is saved the restrictive way, its switch is greyed out for them.
When a change is refused #
A teammate should rarely see a refusal, since Settings holds these controls back. One comes from a page opened before an update, say, or from a script. Nothing in that save is stored, and the message names each setting as Settings does:
| The message | What it means |
|---|---|
| Forbidden: Guardian Vetting is admin-only. | A setting in the table above; ask an administrator at the machine to change it |
| Forbidden: only an administrator can lift these Remote Control restrictions: Read-only for remote devices. Turning one on is allowed. | Loosening a Remote Control restriction; see Who can change these switches |
| Forbidden: only an administrator can turn on Voice on remote devices. Turning it off is allowed. | See Voice from a paired phone |
| Forbidden: Retrospective Loop needs the HITL edit permission (Settings › Access). | The Retrospective Loop's switch, schedule and token cap need Edit on Sidebar: HITL Button & Modal |
| Forbidden: … is managed by your organisation's policy. | A setting your organisation's policy locks; see How a policy is applied |
| These settings cannot be changed from a remote device. Change them in C5 on the machine itself: Guardian Vetting. | Saved from a paired phone or tablet: change it on the machine itself |
| Strict Path Boundary needs a value: an empty one would turn it off. Nothing was saved. | An empty value (below) |
Sending the value already stored is not a change, and goes through. For Guardian Vetting what counts is whether vetting is on now, wherever that was set: sending the state in force goes through, and any change of it is refused.
A blank never turns a guard off. An empty value for Strict Path Boundary, the Egress Data-Loss Prevention mode or the filesystem autonomy level is refused from everyone, an administrator included; an administrator turns one off by choosing off. A setting nobody has saved yet counts as what C5 uses while it is unset, and one C5 cannot read at that moment counts as a change, so neither lets a teammate's value through as a repeat.
Hard limits #
Settings → Guardrails holds the checks that run no matter what any other setting says. They cover things like blocking private information from being sent out, refusing work when the safety checks themselves cannot be reached, and requiring an approved budget before an agent spends anything.
Guardrails beat permissions. If the two disagree, the guardrail wins.
Warning Guardrails and the permission ladder are your real controls. Settings → Locations is not one of them — it tells C5 which folders to bring documents in from and put results into. It is a convenience setting, not a fence: do not rely on it to keep an agent away from the rest of your disk.
If you need a genuine boundary, run C5 under a user account that only has access to what you are willing to expose.
Your choices follow the tool that actually runs #
Some tools have more than one name. A model may ask for bash when it means C5's
run_command, or create_file when it means write_file. When that happens, C5 looks up
your choices under both names and uses the stricter one. Setting run_command to
No for an agent therefore also stops it asking for bash, and Always Allow given
when it asked for bash is remembered for run_command too, so you are not asked again.
Only C5's own read-only tools are approved unattended for being read-only. A tool
from an add-on (MCP) server is never trusted just because its name sounds harmless:
a third-party tool called search still asks, and still goes through Guardian. The one
exception is C5's own web search and page reading: when you connect one of the search
services C5 knows (such as Brave, DuckDuckGo or SearXNG) and C5 uses it to carry out its
web_search, it stays approved, as native search is.
When C5 cannot read a setting #
If C5 cannot read a safety setting (the database is busy or a value is damaged), it never takes that as permission:
- An agent's tool list that cannot be read refuses that call rather than allowing it.
- A damaged list of denied tools refuses every call for that agent until the list is saved again.
- A damaged Master mode setting makes every tool ask, and C5 records a system event so you know to re-save it.
- A data-loss-prevention setting C5 does not recognise is treated as Block. Only an explicit Off turns scanning off.
Autonomy thresholds apply to every run #
The per-work-type thresholds under Settings → Autonomy apply to every step C5 runs, not only to agents that claim work from outside. When a step matches several kinds of work (say, both "billing" and "running code"), every matching rule is checked and the strictest outcome wins. An agent below a rule's level is either stopped or told what it may not do, as that rule says. The kind of work is read from the task's tags and the step's own instructions, not from the earlier results it was given. Reviewing and planning steps are never told to hold back this way (a reviewer checking a deployment is not deploying), though a rule set to stop work still stops them, and their tools still follow the ladder. A stopped step keeps its instructions, so once you raise the agent's level and move the card back, it picks up exactly where it was. Tasks involving personal data always run at level 1 with sign-off, and an agent below level 3 that handles personal data is also told to keep it inside the task's workspace, never send it to an outside tool, service or web address, and mask it in what it writes unless the task needs it.
Never hand over credentials #
C5's guardrails watch for private information leaving your machine, but do not rely on software to notice a password. There is no check that recognizes a card number or a security code and stops on your behalf.
So make it a habit:
- Never put a password, card number, or one-time code into a chat message or a task description.
- Never ask an agent to sign in or pay as you. If a job needs that step, do that step yourself and let the agent carry on afterwards.
- Keep keys in Settings, which stores them encrypted and sends them only to the service they belong to — not in your instructions, where they become part of the conversation.
Reviewing what happened #
The Audit view in Ops lists every permission decision — what was asked, what you chose, and when. Worth a look after your first busy week to see whether your settings match how you actually work.
A good starting point #
- For your first week, leave anything that deletes, sends, spends, or controls your screen set to ask you.
- Notice which prompts you approve every single time.
- Loosen just those — set those specific tools to On in Settings → Extensions.
- Leave the rest asking.
You will end up with settings that fit your work instead of someone else's guess.
