Launched this week
Caw
Open source web terminal multiplexer for AI agents
48 followers
Open source web terminal multiplexer for AI agents
48 followers
Caw is a web terminal multiplexer for OpenCode, Claude Code, Codex, Copilot, Antigravity, and Pi. It makes it easy to monitor and interact with multiple agent sessions from anywhere, especially on mobile. Besides the terminal multiplexer itself, it includes: - Built-in file browser and editor - Kanban-like board showing all running agents - Push notifications whenever an agent finishes working or requests your input - Git worktrees on demand to parallelize work - Voice mode
Caw
Multiplexing terminal sessions specifically for AI agent workflows is a smart niche — curious how Caw handles session persistence when an agent's process crashes mid-task, does it auto-reattach or do you lose scrollback? Also wondering if it's local-only or if there's a relay involved, since that matters a lot for anyone running agents against sensitive codebases. Open source is the right call here given how much trust people need to hand a tool that sits between them and their terminal.
Caw
Hi Valeria, thanks for your comment!
The scrollback is only kept in a small in-memory buffer so, on crash, you lose the raw scrollback. What survives is the agent's own internal sesión though, and when you reopen a crashed pane, Caw tells the agent to resume where it left off using each agent's own "--continue" command.
Regarding your other question: fully local, no relay. You self-host the binary. The only outbound call is for push notifications and that doesn't carry your data.
Caw
Hey there! Caw doesn't try to parse raw agent output for hints. Instead, it watches each agent's structured transcript and only fires when the agent hits a specific, known "asking for input" tool call (the one each agent uses internally).
By default, you'll get push notifications for both "finished" and "needs input" states, but this is configurable. For example, if you only care about being pinged when an agent is stuck waiting on you, you can turn off notifications for "finished" and just keep "needs input" on.
On worktrees: honestly, this is still an open problem for me. Right now, if a conflict happens, my workaround is just spinning up another agent and asking it to sort out the conflicts. Do you have any ideas on how this could be handled better? Curious what you'd want to see here.
The kanban-of-running-agents plus push-when-input-needed combo is exactly the missing supervision layer — running several coding agents, the cost isn't compute, it's ME polling terminals to see who's stuck. Question on the git-worktrees-on-demand: when two agents' worktrees both touch the same file and diverge, does Caw surface the conflict before merge time, or do I find out at the end? That's where parallel agent work usually bites. Congrats — starring the repo.
Caw
Hi @nitish_garg4, thanks for your comment.
Caw does not surface the conflict before merge time. For this I personally have a skill so I get to fix merge conflicts with one simple slash command, however I am open to any idea regarding this.