Instructions and Discord @mentions

September 14–20 brought standing Instructions for the agent, a Discord integration, and conversations that stay in sync across your computers and phone. Plans also got a cleaner layout, and the iOS app now sends push notifications.

Instructions

You can now give the agent standing guidance that applies to every conversation, much like a CLAUDE.md or AGENTS.md file. Add short markdown instructions in Settings → Instructions:

• Personal instructions apply only to your own conversations. Any member can add them.
• Team instructions apply to everyone in the team. Only administrators can edit them.

When the two conflict, your personal instruction wins unless the team instruction is a hard policy. Claude Code and Codex agents can also read and edit Instructions for you when you ask.

Discord

Invite the Nuphos bot to your Discord server and start a conversation by mentioning it. Nuphos replies in a thread. Within that thread you can keep talking without mentioning it again, and it keeps the surrounding discussion as context.

• A channel stays silent until someone runs /nuphos enable. That person needs Discord's Manage Channels permission and must be a Nuphos workspace administrator.
• Use /nuphos full-access or /nuphos auto to choose the permission mode for your own turns in a thread.
• Approval buttons work directly in Discord.
• Administrators can disconnect Discord from the connector page, or switch to another server.

Conversations that follow you across devices

Conversations, unread markers and running replies now stay consistent across your computers and phone.

• Reading a conversation on one device clears its unread dot everywhere.
• The conversation list shows replies that are still running, including replies started from your phone, Slack or a Trigger.
• When a turn has ended but background tasks are still running, the conversation shows an orange dot instead of a spinner.
• Local tools such as local_exec and port forwards now run only on the device that started the turn, never on two machines at once.

Plans

Plan details now open as full workspace pages instead of a sidebar. In finished turns, plan cards move below the final response and start collapsed. The approve and reject buttons stay visible. While a plan waits for approval, the agent is blocked only from running the plan's own commands. It can keep doing unrelated work.

More improvements

• The agent can look up and read your past conversations when you refer to earlier work.
• The agent picker and agent settings show how much of your Claude Code or Codex quota is left, and when it resets.
• Starred chats now appear in a Pinned list above Chats.
• Every connector now offers Add another. Slack, Lark and LINE WORKS can be reinstalled, reconfigured or disconnected from their settings.
• Links to GitHub, Grafana, the AWS and GCP consoles, and Cloudflare open inside the app when the resource is connected.
• Grafana dashboards render pie charts, bar gauges and markdown text panels, and fit narrow panes.
• Triggers run in Full Access with all of the team's current credentials by default, including connectors added after the Trigger was created.
• A credential that is no longer available is now dropped from the conversation without an error.
• Agent settings chart disk usage and sessions per agent.

Mobile

• Push notifications when a turn finishes or fails, or when an approval or Plan is waiting for you. Reading a conversation on another device clears the badge.
• Tool calls appear as grouped tool runs, like on Desktop.
• Larger, more readable transcript text that follows Dynamic Type.
• Long-press a message to copy it, or choose Select text.
• A new Connectors page lists your team's connectors.

Removed

• iMessage and LINE WORKS channels have been removed. To talk to Nuphos from chat, use Slack, Lark, or the new Discord integration.

Fixes

• Long-running agent turns, background tasks and Codex tool calls recover more reliably after a restart or a dropped connection.
• Messages sent mid-turn reach the agent reliably.
• Tool calls and narration stay in the order the agent produced them, and interrupted turns fold like completed ones.

Available in Desktop v0.58.0 and later.