Sophon 1.11: Proactive Agents, Real-Time Voice, and a Mobile Redesign

Sophon 1.11 is built around one idea: your assistant stops waiting for you to ask. Agents now run standing checklists on a schedule, transcribe your speech as you talk, understand whatever media you send them, and behave properly in busy group chats — all without stepping outside the approval guardrails you set.

Heartbeat: Agents That Check In On Their Own

Write a checklist — triage my inbox, watch the build, prep tomorrow's meetings — and your agent works through it on a schedule, unattended. The checklist is a simple contract: when a run finds nothing worth doing, the agent replies HEARTBEAT_OK and the session is archived automatically, so noop runs never clutter your session list or ping your phone. When a run takes real action, you get a notification.

Unattended doesn't mean unguarded. Heartbeat runs can auto-approve tool calls up to Medium risk — High and Critical actions always page you, and reject on timeout if you don't answer. Read the full guide in Heartbeat.

Speech That Appears as You Talk

Sophon now streams microphone audio to the server and transcribes it in real time. Interim transcripts appear on screen as you speak and finalize when you pause — hands-free with automatic pause detection, or push-to-talk when you want control. And the Dashboard now listens through the same server-side pipeline as every other client — no browser speech APIs required.

Pick your transcription provider: Deepgram, OpenAI, Azure, Google Cloud, ElevenLabs, or self-hosted Sophon Managed Speech. Details in Voice.

A Mobile App That Feels Native

Sophon Mobile has been rebuilt on Material 3 with three themes — Daylight, Indigo Core, and AMOLED — plus Material You dynamic color on Android 12+, edge-to-edge layout, and predictive back gestures.

Under the new surface is serious reliability work: per-session stream buffers mean concurrent conversations never cross wires, mid-task reconnection reconciles partial streams into the full message, and pending approvals are recovered when you come back online. See Mobile docs or the mobile overview.

Know Your Models — and What They Cost

The new model catalog ships capability and pricing data for 20+ bundled models across Anthropic, OpenAI, Google, and Ollama: context windows, vision, function calling, reasoning support, and flat per-million-token costs, all in one matrix. Hide models you don't want offered, and switch models per session — from the Dashboard chat picker or with /provider in the CLI. Explore it in the Model Catalog.

Group Chats Done Right

Dropping an agent into a busy Slack workspace or Discord server used to be all-or-nothing. Now it's policy-driven: per-channel controls decide when the agent responds — mention-only (the default), an allowlist of approved groups, all messages, or disabled — with mention, reply-to-bot, quote, and thread triggers detected across six platforms.

The conversation mechanics got smarter too. Rapid-fire messages coalesce into a single agent turn instead of triggering separate replies. Quoted messages are captured and shown to the agent so it knows what you're referring to. And threads — Slack threads, Discord forum posts, Telegram topics — each get their own isolated session with automatic expiry. Read Group Chats and Threads.

Sophon Sees What You Send

Send an image, a voice note, or a video to any channel and Sophon understands it. Audio is transcribed automatically, images are captioned for models without vision, and videos are digested asynchronously with transcript extraction and key-frame captioning — so any model can reason over any media. More in Media Understanding.

It works the other direction too: ask for a slide deck and Sophon builds a complete themed .pptx in one call — 10 layouts, 6 themes, speaker notes, and follow-up editing operations to refine the deck. See Presentations.

A Sharper CLI

CLI v2 is now the default. The input zone stays pinned at the bottom while messages stream above it, slash commands and @ file mentions get inline pickers right in the prompt, and a live status band shows active tasks and in-flight subagents with elapsed times. New chat commands round it out: /provider (a per-session model override, by design), /thinking, /auto-approve, and /canvas — your thinking and auto-approve settings persist across sessions. Four terminal themes included.

Under the Hood

Realtime events are now session-scoped: badges, approvals, and task updates target only the session they belong to, so working across Dashboard, Mobile, CLI, and Desktop no longer produces cross-device noise — and on shared gateways, every live event is tenant-scoped. The release also lands a round of production hardening: automatic credential migration into the vault, channel configuration rollback on failure, SSRF protections, and configurable rate limits.

Get Sophon 1.11

The full list of changes is in the changelog, and every feature above has a guide in the docs. Update your deployment and let your agents take the first step.