AI

Coding agents and the Agent hub

See what your AI coding agents are doing across every terminal, answer them without reading raw output, and run them at scale.

Synapse can run any command-line coding agent - Claude Code, Codex, Gemini CLI, Grok CLI, Aider, OpenCode, Cursor Agent and others - in a normal terminal pane. The Agent hub and Agent workflows services build on that to watch, answer and coordinate them. Switch either on in Services.

Agent hub

With Agent hub on, Synapse watches the visible screen of every terminal running a recognised agent and works out its status: working, idle, waiting for your input or approval, hit a usage limit, errored, or exited.

• A "needs you" badge and, if you want, a desktop notification and sound appear when an agent needs input, approval, hits an error or a usage limit - configured in Settings > AI > Agent hub, including a "quiet while I'm looking" option that skips the alert if you're already looking at that pane.
• The Agent dashboard (Ctrl+Shift+A) lists every agent across every host and session, with its status, last activity and current prompt, and lets you answer a waiting question without leaving the dashboard.
• Focus view replaces a pane's raw terminal output with a calm summary of what the agent is doing and a clear box for its current question - "Show terminal" is always one click away back to the real thing.
• Resume scans a host in the background (nothing is typed into your terminal) for agent sessions already running inside tmux, which you can reattach to, and recent Claude Code or Codex conversations you can resume from where they left off.
• Transcripts keeps a readable, searchable record of your conversations with agents, with secrets hidden, stored on this computer - turn it off in Settings if you'd rather nothing is saved.
• The prompt composer (Ctrl+G, while an agent's pane has focus) opens a proper text box for writing a prompt, with buttons to insert your current terminal selection or a file path, and a history of recent prompts to reuse.

Structured mode (ACP)

Some agents - Claude Code, Codex, Gemini CLI and OpenCode - can run through the Agent Client Protocol instead of a plain terminal. When Agent hub is on and the matching adapter is installed on the host, launching one of these agents offers a Structured mode (ACP) toggle. It replaces the terminal screen with a proper chat: your messages and the agent's, collapsible thinking, a card for every tool action with its real status, a live plan with progress, and explicit Allow/Deny buttons for approvals - all driven by the protocol's own state rather than guessed from on-screen text. You can flip a pane between structured and plain-terminal view at any time.

Agent workflows

Switch on Agent workflows to run and review agent work at a larger scale:

• Parallel agents in git worktrees - start a new worktree and branch for a task, with an agent launched straight into it, optionally inside tmux so it keeps running if the connection drops.
• Review changes - a full diff view of a worktree's changes with line comments, an overall comment box, and buttons to commit, push or open a pull request (using the GitHub CLI on the host).
• Task board - a Kanban board that tracks a task from To do through to Done as the agent working on it makes progress.
• Scheduled runs - a repeating schedule (daily, weekdays, or every few hours) that launches an agent with a saved prompt. Schedules only run while Synapse is open; for one that runs without it, add it to the host's crontab instead.
• Plan review - read an agent's current plan, comment on individual steps, and approve it or send back your comments as a reply.
• Relay across all open sessions (opt-in) - see Relay assistant.

Agents on your phone

With the Sync between devices service on, Settings > AI > Agent hub has a "Show agents on my phone" toggle. Turning it on sends a short summary of your agents - name, host, status and any question waiting for you, secrets removed - to your paired phone, end-to-end encrypted through your sync mailbox, so you can answer a waiting agent without being at your computer. No terminal output is included.