Channels & Messaging

Connection happens in two stages. First, a platform admin sets up the Slack app under Settings → Integrations → Slack Integration, following the checklist shown inline: under OAuth Credentials, enter the app's Client ID, Client Secret and Signing Secret and click Save Credentials; under Transport Connection, paste the app-level token (xapp-...) and click Connect to start Socket Mode (the badge reads Socket Mode, Webhook or Disconnected); then click Install to Workspace (Reinstall to Workspace once a workspace is connected) to complete OAuth and obtain the bot token. The client secret, signing secret and app token are stored encrypted at rest. Second, each agent is bound to Slack from its own detail page: open the Sharing tab, click Configure on the Slack row, and click Create Channelto create a dedicated channel bound to that agent — a public link's Connect Slack button is a shortcut into the same binding (see the Sharing, Access & Monetization FAQ). Only one Slack workspace can be connected per Trinity instance. See Slack Integration.

Socket Mode is the default transport: Trinity opens an outbound WebSocket connection to Slack, so your instance needs no public URL — it works even on a laptop behind NAT. Webhook mode is the HTTP fallback for production environments where Slack posts events to your server, which requires a publicly accessible URL. If you can use Socket Mode, it is the simpler option. See Slack Integration.

Yes. Each agent gets its own dedicated Slack channel via Create Channelin the agent's Sharing tab, and @mentions in that channel route to the bound agent. Direct messages to the bot go to a single workspace-wide DM-default agent instead. Replies post with each agent's own name and avatar, so multiple agents in one workspace stay visually distinct. See Slack Integration.

By default all agents share one workspace bot and are told apart by their per-message name and avatar. Giving an agent its own dedicated Slack bot— one people can DM directly and @mentionby name, alongside other agents in the same channel — is an enterprise-tier capability, configured from the agent's Sharing tab. See Slack Integration.

Yes. Proactively posting to a Slack channel requires a per-channel consent toggle, separate from the per-recipient consent for DMs — a newly bound channel denies proactive posts until you enable it in the Slack channel panel. Replying inline to someone's message never needs consent. See Slack Integration.

DMs are routed to the workspace's DM-default agent. The first agent bound to the workspace becomes the DM default automatically, but you can reassign it anytime: open the target agent's Sharing tab → Slack section and click Set as DM Default— the change takes effect immediately, no restart needed. Only one agent per workspace can be the DM default, and you cannot unbind the current DM-default agent while other agents are still bound; reassign the default first. See Slack Integration.

That is thread tracking working as designed. Once an agent has been @mentioned in a channel (or a thread starts from one of the bot's own messages), replies in that thread route to the same agent automatically, so people can hold a conversation without re-mentioning the bot on every message. Thread tracking only applies to threads started by a bot message or an @mention — the bot does not join arbitrary threads. See Slack Integration.

Trinity posts each reply with the agent's name and its Trinity avatar as the per-message bot icon, using Slack's chat:write.customizescope. This applies to both replies and proactive messages, so in a multi-agent workspace you can always tell which agent is speaking. The agent also sees who is talking and where — channel messages arrive with a prefix like [Channel: #engineering] / [From: John Smith (@johndoe)], while DMs stay clean. See Slack Integration.

Create a bot first: message @BotFather on Telegram, send /newbot, choose a name and a username ending in bot, and copy the bot token. Then open the agent's Sharing tab, click Configure on the Telegram row, paste the token, and click Connect Bot — Trinity validates the token and registers the webhook. Each agent needs its own dedicated bot; bots cannot be shared across agents. If the status shows the yellow "Bot connected but webhook not registered", set Public URL under Settings → Generalto your instance's public address — the webhook registers automatically once saved. See Telegram Integration.

Two settings control this. In Trinity, each group has a trigger mode: Mention only(the default — the bot responds to @mentions and replies) or All messages. For "all messages" to work, you must also disable Telegram's Privacy Mode via @BotFather (/setprivacy→ Disable), because with Privacy Mode on, Telegram simply never delivers non-mention group messages to the bot. Critically, if the bot is already in the group, you must remove and re-add it after changing Privacy Mode — Telegram only applies the change to newly joined groups. See Telegram Integration.

Yes. Voice notes are automatically transcribed with Gemini and delivered to the agent as text prefixed with a 🎙️ emoji — users just send voice notes normally. Limits: 5 minutes duration and 10 MB file size, and the platform needs a Gemini key (next question). If transcription fails or no key is configured, the agent receives a placeholder such as [Voice message received — transcription failed] instead, so the conversation still progresses. See Telegram Integration.

gemini-3.5-flash by default; set GEMINI_TRANSCRIPTION_MODELon the backend to use another. The key is the platform Gemini key — saved from the browser under Settings → Integrations → API Keys (or the Other keys step of first-run setup), with GEMINI_API_KEY in the server's .env as the fallback when nothing is saved in Settings. It must start with AIza, and it is the same key that powers voice conversations and generated avatars, so adding it once switches all three on. See Telegram Integration and Platform Keys.

Yes. Replies are converted to Telegram formatting — bold, italic, strikethrough, inline code and fenced code blocks, headings, links, lists, spoilers, and blockquotes (long ones collapse as expandable quotes); tables become preformatted text. The conversion escapes first, so a literal <, >, or &in a code snippet, an HTML tag, or a comparison survives instead of breaking the message. A reply longer than Telegram's 4,096-character limit is split into several messages, and formatting that straddles a cut is closed at the end of one chunk and reopened in the next, so no chunk arrives mangled. Proactive messages and completion reports use the same rendering. See Telegram Integration.

Trinity connects to WhatsApp through Twilio — you bring your own Twilio account. Open the agent's Sharing tab, click Configure on the WhatsApp row, and enter your Twilio Account SID, Auth Token (stored encrypted), and WhatsApp sender number in the form whatsapp:+15551234567. After connecting, copy the generated webhook URL into your Twilio console so inbound messages reach Trinity. You also need Settings → General → Public URL set to your public domain, and if you route traffic through a Cloudflare Tunnel, an ingress rule sending /api/whatsapp/webhook/* to the backend service (not the frontend). See WhatsApp Integration.

Use the sandbox for development: it is free and uses Twilio's shared sender whatsapp:+14155238886, but every tester must first opt in by sending join <your-sandbox-keyword>to that number. For production, register a dedicated sender in the Twilio console — this requires linking a Meta Business Manager account, and display-name approval takes 24–48 hours. Trinity auto-detects the sandbox from the well-known sender number. See WhatsApp Integration.

No. Twilio's WhatsApp API supports direct messages only, so WhatsApp groups are not available — this is a Twilio/WhatsApp platform limitation, not a Trinity setting. If you need group conversations, Telegram and Slack both support them. See WhatsApp Integration.

Users send /login user@example.com to the bot, receive a 6-digit code at that address, and reply /login 123456 to complete verification. /whoami shows the currently verified email and /logoutclears it. On Slack no command is needed — the workspace OAuth already provides each user's email. Verification matters for agents with a restricted access policy: unverified users are prompted to log in before the agent responds. See Access Control.

Create a public link: open the agent's Sharing tab, go to Distribution → Public links, and click Create Public Link. You can configure email verification on or off, rate limits, and a custom welcome message, then share the generated URL. Recipients open it and chat immediately — no login required. See Public Links.

Yes, if email verification is enabled on the link. Verified sessions persist across page refreshes and return visits, and the agent maintains per-user memory scoped to that agent and email, updated by background summarization every 5 messages. Logged-in Trinity users additionally get a history dropdown to resume any past session. Anonymous sessions (verification off) have no cross-session continuity. See Public Links.

Yes. The Public chat modelsetting on the agent's Sharing tab overrides the model for public-facing surfaces only: public links, Slack/Telegram/WhatsApp channels, and paid chat. Your own authenticated chats and scheduled runs are unaffected, and the default option inherits the platform-wide model. This is useful for putting a cheaper or faster model in front of outside audiences. See Agent Sharing & Access.

Yes. The Additional instructions — public & channel chats only field on the Sharing tab (up to 4,000 characters) is injected into the agent's system prompt for outside audiences only: public links, Slack/Telegram/WhatsApp, and paid chat. It never affects your own authenticated chats, scheduled runs, loops, or agent-to-agent calls, so you can set tone, disclosure rules, or scope limits for strangers without touching the agent's core instructions. Leave it empty to disable. See Agent Sharing & Access.

Yes, one reply at a time. Voice is a per-message choice the agent makes during a channel turn by calling the send_voice_reply tool — replies stay text unless it asks, and there is no always-on voice mode. Delivery uses each channel's native format: an OGG/Opus voice note on Telegram, an inline MP3 clip in the Slack thread, and an OGG voice note via Twilio media on WhatsApp. Setup is one platform key plus one per-agent configuration: an admin stores the ElevenLabs key under Settings → General → Voice (ElevenLabs) (encrypted, never echoed back, with an optional platform default voice), then the owner turns Voice replies on for the agent on its Settings tab and picks a voice (or relies on the platform default), and allows it per channel with the Voice repliestoggle inside each channel's dialog on the Sharing tab — a channel left off stays text-only. Voice replies do not cover the web chat or public links. See Voice Replies.

Because voice is opt-in per message and fails soft — a voice problem never loses a message. The most common cause is simply that the agent did not call send_voice_replyfor that reply. Otherwise the reply falls back to text when voice replies are off for the agent, when that channel's allow flag is off, when the platform has no ElevenLabs key, when the reply exceeds the character cap (TTS_MAX_CHARS, default 1,500), or when synthesis, transcoding, hosting, or upload fails for any reason. If it happens consistently, check the key in Settings, the agent-level toggle, and the per-channel flag. See Voice Replies.

Yes — agents can dial a number through Twilio and hold a live, interruptible spoken conversation powered by Gemini Live. VoIP is off by default: the platform needs VOIP_ENABLED=true plus a Gemini key, the Public URLset in Settings, and a publicly reachable deployment (Twilio must open a WebSocket to your instance). The agent owner then configures a per-agent Twilio voice binding in the Sharing tab's Voice calls row, and calls are placed via the call_userMCP tool or the API. This release is outbound only — agents do not answer incoming calls. See VoIP Telephony.

Calls are rate-limited to 5 per owner and destination number per 60 seconds, capped by a per-agent daily call limit (default 50, overridable on the binding), and hard-capped at 10 minutes each. Calls bill to the agent owner's own Twilio account. After the call ends, the full transcript is saved to the agent's chat history with source voice and, by default, dispatched back to the agent as a task so it can follow up on what was discussed; unanswered calls skip this. See VoIP Telephony.

The verified email is the identity across every channel: a user verified on Telegram, WhatsApp, or a public link, or identified via Slack workspace OAuth, is the same person to Trinity everywhere. That email is checked against the agent's access policy — open access lets anyone chat, while restricted agents admit only the owner, admins, and emails on the shared-access list, with everyone else generating a pending access request. Approving a user once admits them on all channels. The Sharing tab also shows a client roster of external channel users who have messaged the agent. For the approval flow and policy details, see Access Control.

Yes. An admin can tune the anti-spam caps on agent-initiated ("proactive") sends under Settings → General → Proactive message limits: Slack per-channel and per-agent, Telegram per-group and per-agent, and proactive direct messages per recipient — each a per-hour limit. The shipped defaults are 10/hour per channel/group/recipient and 100/hour per agent, so nothing changes until you raise them. Setting a value to 0makes that cap unlimited (the guardrail is disabled and the save warns you). Changes take effect immediately, with no restart. This is the setting to raise for a legitimate high-volume agent — for example, one that posts a Slack message for each inbound support request.

Note: these caps apply only to messages the agent starts. Replies to inbound messages (a DM, an @mention, or a thread reply through a channel) are never limited by them. See Slack Integration.

Yes, three layers, all fail-soft and on by default per bot binding. A 👀 reaction lands on your message the moment the agent picks it up (and is cleared when the turn ends), the standard typing indicator fires, and past 30 seconds the bot posts a quiet “⏳ Working on it — N min elapsed” note (sent without a notification) that it edits about once a minute and deletes when the turn finishes. Turn it off in the Telegram dialog on the agent's Sharing tab. In groups it only fires on @mentions and replies — a bot in observe mode stays silent, since a visible reaction would reveal that it is listening. See Telegram Integration.

Yes. When work that started from Telegram finishes later — because it was delegated or run in the background — the result is posted back into the originating chat, threaded to your original message and sent by the same bot you were talking to. Failures report honestly rather than vanishing. An ordinary chat turn already answers inline, so it is never double-posted. Group chats have a per-group consent flag, allow_proactive(on by default), that suppresses these when turned off; direct messages are consented by construction. Slack behaves the same way, gated on the channel's proactive-consent toggle — a channel that denies proactive posts gets no completion note. The same report-back also reaches the Workspace chat a job was started from. See Telegram Integration, Slack Integration, and Event Subscriptions.

Because channel replies run with a restricted tool set: by default only WebSearch and WebFetch, with Readadded automatically when the message carries readable files, and a 120-second execution timeout. Those are platform settings — channel_allowed_tools for the tools, channel_rate_limit_max and channel_rate_limit_windowfor the per-Slack-user rate limit (30 messages per 60 seconds by default) — so an admin can widen the list for an agent that genuinely needs more. Rate limits are per Slack user, not per agent. See Slack Integration.

Most likely the group's proactive consent is off. Every connected group carries an allow_proactiveflag — on by default — that governs whether the bot may post there without being asked: agent-initiated messages via send_group_message(channel_type: "telegram") and the completion report-back for delegated work. Turn it back on with PUT /api/agents/{name}/telegram/groups/{id}(a human-only route). If the flag is on, check that the group hasn't been removed (deactivated) in the Sharing tab's Telegram Groups section and that the agent's proactive caps under Settings → General → Proactive message limits haven't been hit. Direct messages never need consent — the user started the conversation. See Telegram Integration.

Inbound Slack files are downloaded through a host allow-list: the authenticated first request may only reach Slack's own API host, and redirects are followed only to Slack's file and edge-CDN hosts, over HTTPS, with each hop re-checked and a bounded hop count. A download to any other host is refused — and logged as an error, so a refusal is visible in the platform log rather than silently dropped. Check platform.json for the refusal before assuming the message never arrived. See Slack Integration.