Trust and security
AI agent rules
You can connect an AI assistant to Harmny so it can work with your tasks. This page lists the rules it works under: what it can read, what it can change, who approves it, how its changes are recorded and how to turn it off.
Setup steps are in how to connect Claude. Limits, scopes and the tool list are on the MCP server page.
Last updated
What an agent can read
An agent reads what the person who approved it can read. A connection approved by an admin can read tasks across the workspace, including private spaces and client-visible work. It can never change them.
If you want a narrower view, approve the connection from an account with narrower access. The agent then reads what that account can read in your tasks.
Career frameworks, career paths, people summaries (job title, level, published ratings and active goals) and the organization overview need more. Only an admin can give an agent those. A connection approved by someone who is not an admin can read projects and tasks only.
An agent works inside your workspace and nowhere else. Another organization's data is out of its reach.
If your organization has turned off agent access to goals and tasks, agents cannot read or change tasks.
What an agent can change
Nothing, until an admin of your workspace turns task changes on and a person approves them on the connection. After that, on internal tasks, an agent can:
- create a task (tasks an agent creates are internal);
- update a task's title, status, priority and dates;
- add a description to a task that has none;
- add a person to a task;
- comment on a task.
A status change may wait for approval, the same way it would for a person, when your workflow asks for one.
What an agent can never change
These rules hold for every role, admins included. Harmny checks them on its own servers for every change, whatever the agent asks for.
An agent cannot delete anything. And it can never change:
- Client-visible tasks, and any task under one.
- Client-portal work, and anything under it.
- Tasks in a private project or a private space.
- A description a task already has (the agent can add a comment instead).
- Another organization's data.
Who can connect an app
Admins, and people an admin has given the Manage Integrations permission in an access group. Only admins can create API keys. Someone who is not an admin can give an agent projects and tasks only.
Before an app connects, the person approving it sees a consent page in Harmny. It shows the app's name, the site it will send them back to, what the app will be able to do, and the workspace. They choose Approve or Deny.
Next to the app's name the page says “Unverified app”. The name is whatever the app registered with. It is not proof of who made the app. Only approve a connection you started from that app yourself.
How every change is recorded
Comments and activity an agent writes show the name of the person who approved it, then “via” and the name of the app or the API key. For example: Maria Chen via Weekly Planner. An app's name is marked as unverified.
Every tool call an agent makes also goes into your workspace audit log. Admins can see these entries. Each entry names the person who approved the connection and the app or key. Every change gets its own entry. Repeated reads are grouped by the hour. The log records which action ran and whether it worked, not the content sent.
Your automations
Your automations still run when an agent changes a task. Two actions are skipped for agent changes: emails to outside addresses and webhooks. Team members still get their emails. Skipped webhooks are noted on the automation run.
An automation you set up can still change tasks the agent itself cannot.
How to disconnect
- Sign-in connections: an admin, or someone with the Manage Integrations permission, opens Settings, Connected apps. It lists every app connected to the workspace, who approved it and when it was last used. Choose Disconnect.
- API keys: created and revoked by admins in Settings, Intelligence, MCP API Keys, on the plans that include it.
A key also stops working when the admin who created it is deactivated or is no longer an admin, and when it reaches its expiry date. A sign-in connection stops when the person who approved it is deactivated or loses the access that let them approve it.
Both take effect on the agent's next call.
Turning task changes on
Task changes are off until an admin of your workspace turns them on in Settings. An admin does this in Settings, Intelligence, with the switch "Let agents create and change tasks", and confirms it. A second admin switch, "Let agents take delegated work", needs the first. Turning task changes off stops connected agents on their next call and turns delegation off.
Once task changes are on, a person still has to approve task changes on each connection, or an admin has to create an API key that allows them.
In rare cases, such as abuse, Harmny can pause agent task changes for a workspace; admins then see that in Settings.
Questions
Can an AI agent delete anything in Harmny?
No. An agent can create, update, assign and comment on internal tasks once an admin of your workspace turns task changes on. There is no way for it to delete a task, a comment or anything else.
Who turns on task changes?
Task changes are off until an admin of your workspace turns them on in Settings.
What can an agent read?
An agent reads what the person who approved it can read. A connection approved by an admin can read tasks across the workspace, including private spaces and client-visible work. It can never change them. If you want a narrower view, approve the connection from an account with narrower access.
Can an agent change client work?
No. An agent can never change client-visible tasks, client-portal work or anything under them, and it can never change a private project or space. This holds for every role, admins included.
Who can connect an AI agent to our workspace?
Admins, and people an admin has given the Manage Integrations permission in an access group. Only admins can create API keys.
What does “Unverified app” mean?
The app name is whatever the app registered with when it connected. It is not proof of who made the app. Only approve a connection you started from that app yourself.
How do I disconnect an agent?
For a sign-in connection, an admin, or someone with the Manage Integrations permission, opens Settings, Connected apps, and chooses Disconnect. API keys are created and revoked by admins in Settings, Intelligence, MCP API Keys, on the plans that include it. Either way, the next call the agent makes is refused.
Want agents to change tasks?
How to turn it on: an admin opens Settings, then Intelligence, and switches on "Let agents create and change tasks".
Sign in to Harmny