Skip to main content
Trinity
Guides/Who Can Do What

Who Can Do What

Three separate answers decide every permission in Trinity. A person's platform role applies on every agent. Their access is set agent by agent. Their relationship to an agent says whom the agent works for, writes to and waits on. This page maps all three, plus what API keys can reach.

Trinity Roles & Permissions Explained: Platform Roles, Agent Access, and Approval Relationships

Nov 2026

Each layer is set separately and answers a different question. What someone can do on a given agent is the three combined: an admin can sit on an agent's roster as a plain viewer, and a Workspace-only user can be the person an agent works for.

1. Platform Roles

Roles stack: admin > creator > operator > user. Each role gets everything below it, and the same checks apply in the UI, the REST API and MCP.

ActionExternal clientVerified email, no accountuserWorkspace onlyoperatorRuns shared agentscreatorBuilds and owns agentsadminRuns the platform
Working with agents
Chat with an agentYes, with a conditionVerified emailYesSharedYesSharedYesOwn + sharedYesAll
Answer asks addressed to themWorkspace InboxYesYesYesYesYes
Open the operator UIDashboard, agent pages, Operations, SettingsNoNoYesYesYes
Run tasks and add schedulesNoNoYesSharedYesOwn + sharedYesAll
Building agents
Create agentsCreator quota: 10 by defaultNoNoNoYesQuotaYesUnlimited
Install a system from a manifestNoNoNoYesYes
Configure, share and delete an agentNoNoNoYesOwnYesAll
Through the API and MCP
Call the operator API and MCP toolsA user gets 403 workspace_onlyNoNoYesYesYes
Create a personal API keyFrom a signed-in session; a key can't create a keyNoNoYesYesYes
Call admin-only endpoints with a personal keyNoNoNoNoYes
Running the platform
Manage users, roles and the email whitelistSigned-in session; not their own roleNoNoNoNoYes
Platform settings, agent quotas, skill sourcesNoNoNoNoYes
Read other people's chats with an agentNoNoNoNoYes
Human-only grantsOps and portal keys, large retention deletions, usage sharingNoNoNoNoYes, with a conditionIn person; no API key
Yes yesYes, with a condition yes, with the condition shownNo no

How someone gets a role. The first admin is created at setup. Anyone else who signs in by email gets the role on their whitelist entry, which defaults to user. An admin changes anyone else's role in Settings → Access → User Management. Lowering a role also lowers the whitelist entry, so signing in again can't bring the old role back. See Roles and Permissions.

An external client signs in but gets no account. They prove an email with a code, and Slack proves it automatically. On a Restricted agent the owner approves that email once. See Access Control.

2. Access to One Agent

A shared person can use an agent but can't control it. An admin can do everything an owner can, on every agent.

On this agentOwnerCreated itShared personAdded in Settings › PeopleAdminEvery agentNot sharedEveryone else
Using it
Open it and chat with itYesYesYesNo
Run its tasksNeeds the operator roleYesYesYesNo
Start and stop the agentNeeds the operator roleYesYesYesNo
Read chat historyYes, with a conditionOwn chatsYes, with a conditionOwn chatsYesEveryone'sNo
See runs and task historyYesYes, with a conditionOnly the ones they startedYesNo
FilesCredential files: .env, .mcp.json, keys, certificatesYesYes, with a conditionNo credential filesYesNo
Container log and activity timelineYesNoYesNo
Add a scheduleNeeds the operator roleYesYesYesNo
Edit or delete a scheduleYes, with a conditionOnes they createdYes, with a conditionOnes they createdYesNo
Controlling it
Enable, disable or run a schedule nowYesNoYesNo
Change settingsAutonomy, resources, timeout, guardrails. Never with the agent's own keyYesNoYesNo
Manage credentialsYesNoYesNo
Share itPeople, external clients, channels, public linksYesNoYesNo
Delete itNobody can delete the system agentYesNoYesNo
Yes yesYes, with a condition yes, with the condition shownNo no

Sharing does not create an account. A shared email signs in to the Workspace with an emailed code. Whether that person also gets the operator UI is their platform role.

Unprompted messages. The agent can always message its owner. Anyone else must first have the proactive-messaging toggle turned on in Settings › People.

The agent is held to the person it is working for. During a shared person's turn the agent's file tools reach only what that person could open, it gets no shell, and it can't set up standing work for them. See Agent Sharing.

3. Relationships to an Agent

A relationship records who a person is to an agent, not their job title. The Primary is the person the agent works for. Approvers sign off on gated work. Collaborators and Viewers follow along.

RelationshipPrimaryThe person the agent works forApproverSigns off on gated workCollaboratorWorks alongside the agentViewerFollows without deciding
How the agent addresses themMessages, reports, asks, voice notesto: primaryto: approverNo value of its ownto: viewer
Approves gated skillsYesBy defaultYes, with a conditionIf the gate names approversNoNo
Approves a gate that names a canon rolerole:<id>Yes, with a conditionIf they hold that roleYes, with a conditionIf they hold that roleNoNo
Receives credential asksThe agent asks for a secret in the chatYesNoNoNo
Gives access to the agentNo. Only sharing gives access. Assigning someone in Settings › People also shares the agent with them; a person who is assigned but not shared shows a warning and a Share with them button.
Where it is setSettings › People → People this agent serves, on instances licensed for it. Non-admins see the section read-only.
When nobody is assignedThe agent's owner stands in for the Primary. Anything addressed to approver or viewer is refused until someone fills that role.
Yes yesYes, with a condition yes, with the condition shownNo no

Assigning people needs an entitlement. Without it there is no People this agent serves section, the owner is the Primary, and approver and viewer are refused. Gates that name the primary contact work on every instance.

An approver must also be able to open the agent. An approval card goes only to people in the role who own the agent, have it shared with them or are admins. See Approvals and Skills and Playbooks.

"Operator" is never a relationship. to: operatorgoes to the instance's operators as a group, with no single person named.

Three Rules People Get Wrong

Rule 1

Assigned is not the same as allowed in

A relationship tells the agent who someone is. Only sharing gives access. If an approver can't open the agent, a gated request is refused with approver_cannot_open_agent. The fix is to share the agent with them.

Rule 2

A role is not the same as a human

An agent's key acts with its owner's role, but it never passes an admin check. Granting power needs a person: roles and new keys need a signed-in browser session, and an agent's settings refuse the agent's own key, so an agent can't raise its own limits.

Rule 3

The layers mix freely

A Workspace-only usercan be an agent's Primary and approve its gated work without ever seeing the operator UI. Raising someone's platform role doesn't change their relationship to any agent, and the reverse holds too.

4. Agents and API Keys

Every API key belongs to a person and can do at most what that person can. Only a browser session, a personal key and the system agent can pass an admin check.

CallerReachesWhat the key can touchAdmin-only endpointsCan it pass an admin check?Created byWho can mint it
Personal keyuserEverything the owner's role allows, except writes that take a signed-in sessionYes, with a conditionIf the owner is an adminAny operator or above, for themselves
Agent's own keyagentIts own agent, the agents it may call, the agents it spawnedNoThe platform, when the agent is created
System agentsystemThe platform's own agent; skips agent permission checksYesThe platform
ConnectorconnectorOne agent's exposed skills and chatNoThe agent's owner
Portal delegateportal_delegateOne route: starts a Workspace session for an end userNoAn admin, in person
Ops, read-onlyopsFleet health, telemetry, roster, executions, usageYes, with a conditionThose reads onlyAn admin, in the browser
Ops, writesops_writeReset a circuit breaker; start, stop an agent; terminate a runYes, with a conditionThose four onlyAn admin, in the browser
Yes yesYes, with a condition yes, with the condition shownNo no

Grants that need a person refuse every key.Changing roles, adding whitelist entries and creating keys take a signed-in session; an admin's own key is refused too.

An ops key follows its owner. It stops working when the admin who created it is demoted or suspended, so create it under a service account. See Authentication → MCP key scopes.

See Also

•Roles and Permissions — how roles are assigned, changed and enforced
•Agent Sharing — sharing an agent with people and external clients
•Access Control — Restricted and Open agents, email verification, access requests
•Approvals — how asks and gated-skill approvals reach the right person
•Authentication — tokens, key scopes and the human-only gates