Skip to main content
Trinity
Updated Nov 17, 2026main@87e21ff
Guides/Agent Sharing & Access

Agent Sharing & Access

Two owner-only surfaces on the Agent Detail page control who can reach an agent: Settings › People (the former Access tab) manages the people the agent is shared with (they use it in the Workspace), and the Channels tab manages external clients (outside users reaching the agent through Slack, Telegram, WhatsApp, voice calls, or public links).

Concepts

•Operator — A Trinity platform account with role operator or above. Operators log into the Trinity UI and get interact-level access to agents shared with them. A platform account with role user works in the Workspace only — see Roles and Permissions.
•External client — Someone without a Trinity account who reaches the agent through a channel (Slack, Telegram, WhatsApp, a phone call) or a public link. Identified by verified email.
•Workspace roster — Every email an agent is shared with can sign in to the Workspace and chat with it there. Sharing an agent to an email is that person's Workspace account; unsharing removes the agent from their roster.
•Access policy — The per-agent Restricted/Open switch controlling whether unknown verified users may chat. See Access Control for the cross-channel verification model.

Settings › People — Trinity Operators

Settings › People lists the people this agent is shared with. It is the Access tab moved into the Settings home, unchanged.

1

Open the agent detail page, click the Settings tab and pick People in the rail (owner only; ?tab=settings&section=people, and an old ?tab=access link lands there too).

2

Enter an email and click Add person. As you type, the field suggests people who already have an account on this instance, by email, username or name; pick one, or type any address. The person can sign in to the Workspace with a code emailed to them. Sharing does not create a platform account or add the email to the login whitelist; whether someone also gets the operator UI is their platform role, which an admin sets on the person's row here or in Settings → Access.

3
Each row shows a status badge:
  • • Active — the email resolved to an existing Trinity account (shows username, role, and last-active time, or "never signed in"). An admin may see a role select on the row instead of the role label: the role applies on every agent, every change asks for confirmation, and an admin's own row stays read-only.
  • • Pending — no platform account ("No platform account — uses the Workspace"). Where the row offers Give a platform account, an admin can create the account right away (after a confirmation; as user — Workspace only — unless an earlier allow-list entry set another role), so the person can be assigned to the agent before they first sign in; they sign in at /login. Give accounts to colleagues — a platform account sees the team's view of every agent shared with its address — and keep outside clients on a plain share.
4

Each operator row also carries a proactive-messaging toggle — "allow this operator to receive proactive DMs from the agent." It is off by default; turn it on per operator to let the agent message that person unprompted. This is a per-recipient DM permission (PUT /api/agents/{name}/shares/proactive, body {email, allow_proactive}). Trinity records consent in three states — not asked (every new share), declined and consented — and only consented lets the agent send; the other two are both a no, kept apart so you can see whether anyone ever asked.

5

Click Remove on a row to revoke access.

When you add someone with no account, Trinity emails them an invitation that says who invited them and which agent, and links to the Workspace sign-in. Use Resend inviteon their Pending row to send it again, up to 3 times an hour. If email isn't set up for sign-in on your instance, no invitation goes out and the panel says so, so tell them how to sign in yourself.

The invitation's link is built from FRONTEND_URL (in .env) only. If FRONTEND_URL is empty, the email carries no link and tells the person to use your Trinity sign-in page. The Public URL is never used for invitations, because the public chat address serves chat only and would answer the Workspace with an error. Set FRONTEND_URL whenever you set a Public URL.

On instances licensed for it, Settings › People also shows a People this agent serves section above this list: the agent's primary person, its other stakeholders, and the seat the agent holds or serves. Assigning someone there also shares the agent with them, because an assignment by itself grants no access; a row for someone the agent is not shared with carries a warning and Share with them. Without the entitlement the section is absent, and non-admins see it read-only.

The per-recipient DM toggle above is complemented on the channel side by the per-channel Slack consent — see Slack Integration. Platform-wide proactive-messaging rate caps are admin-configurable in Settings.

Access Levels

LevelPermissions
OwnerFull control — create, delete, configure, share, manage credentials and schedules.
Shared operatorInteract only — chat, run tasks, start and stop the agent, view files (except credential files) and the executions they started. Cannot modify config, credentials, or permissions.
AdminFull access to all agents regardless of ownership.

What someone you share an agent with can do

•Chat with the agent and run its tasks. The agent keeps using its credentials while it works for them.
•Open, preview and download the agent's files, except credential files.
•They cannot open or download credential files: .env, .mcp.json, SSH and cloud credentials, key and certificate files (*.key, *.pem, *.p12, *.pfx) and the agent runtimes' login files. Trying shows a message saying so.
•To see or change a credential, they ask the agent's owner or an admin. Managing credentials (set, inject, import, export) is for the owner or an admin.
•If opening the agent's files shows "This agent needs a restart to apply an update", they ask the agent's owner or an admin to stop and start it in Trinity.
•They see only their own work: the executions, task history, live streams and queued prompts they started. They do not see the owner's turns or anyone else's, the container log, the activity timeline or the agent's full chat history. Runs with no person behind them (Slack, Telegram, voice, public-link and scheduled runs) are not in their list. The Tasks panel says "you see the tasks you started".
•They can stop a run only when it is wholly theirs. Changing the agent's model, a forced git pull and managing a schedule's webhook (unless they created that schedule) stay with the owner or an admin. Starting and stopping the agent stay open to them.

What the agent does for someone you share it with

An agent acts for whoever is chatting with it, so Trinity holds the agent to that person's limits for the length of their turn. Trinity decides this from the execution record, not from anything the person or the agent says:

•The owner, an admin, or work the owner set up (a schedule, reminder, loop, webhook or event subscription the owner created) — the agent can do everything it normally can.
•A person the agent is shared with — the agent's file tools reach only the paths that person could open in the Files tab, so it will not edit its own instructions, skills or startup files for them. It gets no shell. It cannot set up standing work for them (schedules, loops, reminders, event subscriptions), create or deploy agents, or read anyone else's executions. On a Gemini CLI or Codex agent their message is refused before it runs, with a reason that says the owner can switch the agent to Claude Code; those runtimes cannot limit what a turn does.
•A channel or public-link visitor (Slack, Telegram, WhatsApp, a public link, voice, a paid call) — the file tools reach the share paths. Whether the agent may also use the shell, run on Gemini or Codex, and set up standing work depends on the owner's Channel and public visitors may run commands switch under Settings › Guardrails (on by default). See Agent Guardrails.

When the agent refuses something for these reasons it says so, so the person can ask the owner. A turn the platform starts to continue earlier work (for example the agent waking up when an ask it raised is answered, or when a delegated run finishes) is held to the same limits as the turn it continues. If the person who answered an ask could not stand as the asking turn (a shared person answering the owner's ask, say), the wake-up is held to a shared person's limits.

Enforcement notes:

•All API endpoints check ownership or sharing status before granting access.
•Shared users see only their own chat messages. Admins see all messages.
•Deleting an agent cascades to remove all associated sharing records.

Channels Tab — External Clients

The Channelstab is organized around "who outside the team can reach this agent, and through what." Sections, top to bottom:

Agent Channels tab showing channels, public chat controls, and distribution options

Who can chat with this agent?

A single Restricted ↔ Open control:

•🔒 Restricted — only approved people chat. Unknown verified users generate a pending access request you approve or deny right below the switch.
•🌐 Open — anyone with a verified email chats immediately.

This is the same cross-channel policy described in Access Control— approving a request admits that email on every channel at once.

Public chat model

Pick the Claude model used for public-facingconversations only — the public link, Slack/Telegram/WhatsApp channels, and paid chat. Your own authenticated chats and scheduled runs are unaffected. The default option ("Use platform default") inherits the platform-wide model; selecting a model overrides it for this agent's public surfaces.

Additional instructions — public & channel chats only

A free-text field (up to 4,000 characters) injected into the agent's system prompt for outside audiences only: public links, Slack/Telegram/WhatsApp, and paid chat. It does not affect your own authenticated chats, scheduled runs, loops, or agent-to-agent calls. Use it to set tone, disclosure rules, or scope limits for strangers without touching the agent's core instructions. Leave empty to disable. (Voice/VoIP has its own prompt — see Voice Chat.)

Channels

Compact status rows for Slack, Telegram, WhatsApp, and Voice calls (when VoIP is enabled platform-wide). Each row shows a connection status dot and a Configure (not connected) or Manage(connected) button that opens the channel's full configuration in a dialog. Per-channel setup guides:

Each channel dialog also carries that channel's voice-replies allow flag — replies are text by default and the agent opts individual replies into voice per message. See Voice Replies.

Portal clients

A list of the people you've shared this agent with, as Workspace (client portal) sign-ins: each row shows when they were last seen, whether they are blocked, and when their sessions were last ended. Refresh reloads it.

•Log out — ends this person's live Workspace sessions everywhere. One sign-in covers every agent shared with them, so there is no per-agent session to end. They can sign straight back in. Any owner of an agent shared with them can do this.
•Block / Unblock — admins only. A block keeps the person out of the whole platform until an admin lifts it, and ends their sessions; their chat history and memory are kept. To remove someone from this agent only, unshare them instead.

Both actions are recorded in the audit log, and an agent's own API key cannot perform them.

Client roster — who's reaching this agent

A table of external channel users who have messaged the agent, aggregated across Telegram and WhatsApp (more channels to follow): client name, channel, verified email, message count, last-active time, and access, ordered by most recent. The roster is database-sourced, so it renders even when the agent container is stopped.

Each row has its own controls:

•Approve — shown on a client whose access request is pending; the same as approving it under Pending requests.
•Revoke — shown on an approved client; removes their access to this agent. While access is Restricted, their next message asks for approval again. While it is Open they can still chat, so block them to stop that.
•Block / Unblock — stops this client reaching this agent on that channel, whatever the access policy. Their messages are dropped without a reply. Their history and any approval are kept, and unblocking restores them. A client who never verified an email can be blocked too.

Each action affects this agent only, is available to the agent's owner (and admins), and is recorded in the audit log.

Distribution

One collapsible panel for distributing the agent's output rather than granting access:

•Public links — shareable chat URLs; see Public Links.

File sharing — signed download URLs for files the agent publishes — is under Settings › Storage, because it governs what leaves the agent's disk; see Agent Files.

For Agents

EndpointMethodDescription
/api/agents/{name}/sharePOSTShare agent with an email address
/api/agents/{name}/share/{email}DELETERemove a share
/api/agents/{name}/sharesGETList all shares for an agent; each carries proactive_consent (not_asked, declined or consented)
/api/agents/{name}/accessGETRoster of people the agent is shared with (active + pending) for Settings › People, with each row's proactive_consent
/api/agents/{name}/shares/proactiveGET / PUTPer-recipient proactive-DM permission; PUT body {email, allow_proactive} (owner/admin)
/api/agents/{name}/clientsGETExternal client roster with each client's client_id and access state (owner-only)
/api/agents/{name}/clients/blockPOSTBlock one channel client on this agent; body {channel, client_id, reason?} (owner/admin, people only)
/api/agents/{name}/clients/unblockPOSTLift that block; body {channel, client_id} (owner/admin, people only)
/api/agents/{name}/clients/revokePOSTRemove an approved client's access to this agent; body {channel, client_id} (owner/admin, people only)
/api/enterprise/client-portal/agents/{name}/clientsGETPortal clients with last-seen, block and session state (owner/admin; the prefix is historical)
/api/enterprise/client-portal/agents/{name}/clients/{email}/logoutPOSTEnd the client's live Workspace sessions (owner/admin, people only)
/api/enterprise/client-portal/agents/{name}/clients/{email}/blockPOST / DELETEBlock (optional {reason}) or unblock the client platform-wide (admin, people only)
/api/agents/{name}/public-promptGET / PUTPublic/channel-only custom instructions (owner-only, ≤4000 chars; empty clears)
/api/agents/{name}/public-channel-modelGET / PUTPublic-channel model override (owner-only; null clears to platform default)

See Also

•Access Control — cross-channel email verification and access requests
•Workspace — where the people you share an agent with chat with it
•Public Links — shareable chat URLs
•Voice Replies — spoken replies on Slack, Telegram, and WhatsApp
•MCP Server — exposing an agent as a dedicated MCP tool