jmux follows the Unix philosophy which means you compose it with the tools for your workflow.

I use a git worktree per thing I am working on and run claude code in the worktree.

So I built jmux. It turns tmux into a parallel development environment. Every session, every agent, every running process, visible and navigable from one terminal. When Claude Code completes a response, an orange flag appears on that session. Switch to it, review the work, move on. No context switching to a different app. No learning a new workflow. It's still tmux underneath.

The key decision was to build on tmux instead of replacing it. My ~/.tmux.conf still works. My neovim setup still works. My keybindings, plugins, prefix key, etc, all of it carries over. jmux just makes it all work at scale.

What it does:

- Run Claude Code, Codex, aider. Any agent, directly, in parallel

- Attention flags. Know the moment an agent finishes without watching every pane

- Instant switching between sessions (Ctrl-Shift-Up/Down, no prefix key)

- Auto-groups sessions by project and git branch

- Works with your existing tmux config, editor, and tools

- No SDK, no API wrappers. First-party tools, no middleman.

What it doesn't do:

- No built-in editor (I use neovim btw)

- No built-in Git (use yours)

- No agent protocol/wrapper (run them directly in the terminal)

~1800 lines of TypeScript, ~0.3 MB installed. MIT licensed.

Landing page: https://jmux.build

Github: https://github.com/jarredkenny/jmux

Comments URL: https://news.ycombinator.com/item?id=47650233

Points: 10

# Comments: 9

Over the past week I've been watching everyone on Twitter/X build their own productivity OS on top of OpenClaw. Everyone was solving the same problems in slightly different ways. I took the best ideas I saw, combined them with everything I've ever wanted in a multi-agent setup, and landed on Agency.

The basic idea: an orchestrator agent breaks down work into tasks, assigns them to worker agents, and reviews results. Workers claim tasks, write code, run tests, and report back through task comments. You manage everything from a dashboard or CLI, or simply by chatting with the orchestrator event in slack/telegram/etc.

Each agent is an OpenClaw instance (https://openclaw.ai/) with full shell, file I/O, and browser access. Agency adds the coordination layer on top.

Some highlights:

- Skill marketplace — Browse and one-click install skills from public repos (anthropics/skills, obra/superpowers, etc.) or import from any GitHub repo that follows the convention

- Deploy anywhere — Agents run as local Bun subprocesses, Docker containers, or on EC2 with automatic reverse SSH tunnels back to the host API

- Claude Max OAuth — Use your Claude subscription directly instead of an API key

- Knowledge sharing — Agents learn facts during work and share them with the team via a shared knowledge base

- Role system — Configure agent behavior per role with Soul, Identity, Tools, Agents, and Heartbeat prompts

- Single package — bun install -g @jx0/agency gives you the API, dashboard, and CLI

The stack is Hono + Kysely + SQLite on Bun for the API, Next.js static export for the dashboard, all served on a single port. The entire state lives in one SQLite file under .agency/.

Still early — I'm running 4 agents across my macbook, my home server, and a fleet of EC2 instances. It works surprisingly well for parallelizing implementation tasks.

Code: https://github.com/jarredkenny/agency-ai

Comments URL: https://news.ycombinator.com/item?id=46850947

Points: 2

# Comments: 1

After recently finding Beads here on HN, my entire development workflow changed (again). I realized that context is state, I am a developer who knows how to handle state.

Naturally I pieced together a couple of existing Claude Code skills, and wrote a couple of my own, and ended up with a workflow that actually delivers on automating the entire idea to shipping pipeline.

- Brainstorming produces a design doc - Design becomes an implementation plan - Plan converts to a Beads epic with inferred dependencies - Epic executes autonomously with two-stage review per task

I'm sharing it on HN because I suspect others are hitting similar problems and working around them in ad-hoc ways.

Nothing groundbreaking, but it actually works.

https://jx0.ca/solving-agent-context-loss/

Comments URL: https://news.ycombinator.com/item?id=46471286

Points: 2

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=45404007

Points: 10

# Comments: 5

Comments URL: https://news.ycombinator.com/item?id=45248918

Points: 2

# Comments: 0

I use claude code exclusively for the initial version of all new features, then I review and iterate. With the Max plan I can have many of these loops going concurrently in git worktrees. I even built a little script to make the workflow better: http://github.com/jarredkenny/cf

I do wonder what it might take to take Nim towards "main stream" status.

Comments URL: https://news.ycombinator.com/item?id=18982378

Points: 2

# Comments: 0