Junction: 7 AI Coding Agents in One VS Code Sidebar

So you’ve got OpenClaw for TypeScript, Hermes for general-purpose agent work, Goose for CLI-first automation, MiMoCode for MCP-native workflows, OpenCode as a Claude Code clone, and Souveraine for reasoning-heavy tasks. That’s six terminals, six CLIs, six sets of commands, six ways to attach a file.

And that’s not even counting OpenHands.

Then I found Junction — a VS Code extension that pulls all seven into one sidebar panel. 512 stars in 3 days as of this writing. Here’s why that number already makes sense.

What Junction Does

But Junction is basically a universal chat frontend for local AI coding agents. Here’s the thing — you install the VS Code extension, make sure at least one agent backend is running on your machine, and Junction’s sidebar lets you talk to it — or switch between them — without leaving your editor.

And the sidebar auto-detects which backends are running, pulls their available models, and presents them in a dropdown. No config files. No --help to figure out which flag does what.

Quick Start — Under 2 Minutes

So I tested this on my Ryzen 9 workstation with OpenCode and Goose already running in the background:

1. Install the extension — From the VS Code Marketplace. Took maybe 15 seconds.
2. Open the sidebar — Command Palette → Junction: Open Sidebar. Done.
3. Pick a backend — The header dropdown showed OpenCode and Goose automatically. I clicked OpenCode.
4. Start talking — Typed “explain the dependency injection pattern in this project,” dragged a src/di/container.ts file into the chat, and got back a clean markdown explanation with the relevant code highlighted.

And it took me about 90 seconds from “install” to “having a real conversation.”

Then I also tested the model selector — since OpenCode supports multiple models, Junction listed them all (GPT-4, Claude, DeepSeek — whatever the backend exposes). And the reasoning slider is a nice touch: turn it up for complex tasks, down for quick questions. Honestly, it just works.

What surprised me most was the tool call rendering. When OpenCode ran a filesystem search, Junction rendered the results as inline diffs and collapsible cards — not raw JSON or terminal scrollback. And that’s a quality-of-life improvement I didn’t know I needed until I had it.

Where Junction Falls Short

Now, it’s honest to say this project is 3 days old — and it shows in a few places:

• Requires agents pre-running — Junction doesn’t start agents for you. You still need OpenCode, Goose, or whatever backend running in a terminal or as a daemon. That’s fine for power users, but less polished than a fully-integrated experience like Continue.dev.
• VS Code only — If you work in JetBrains, Zed, or Neovim, you’re out of luck for now. The developer mentioned web support in the roadmap, but nothing concrete yet.
• 6 open issues — mostly feature requests and minor display bugs. The core chat flow works, but the tool call rendering can lag on very long responses.

Still — for an extension that’s been public for 72 hours, the stability surprised me.

Junction in the Agent Ecosystem

So Junction fits neatly into the broader agent stack we’ve been building here. If you’re already running Hermes or Pi Agent Harness as your day-to-day coding agents, Junction eliminates the context-switch tax of jumping between multiple terminals.

But it also pairs well with the idea of running agent backends on a remote VPS — if you’ve got a VPS setup with Hermes or OpenClaw running persistently, Junction becomes your local VS Code frontend to that remote backend. Same chat interface, zero terminal juggling.

The Bottom Line on Junction

So here’s the bottom line: Junction solves a real problem I’ve been feeling — too many agent CLIs, one desk. It’s not a replacement for any individual agent; it’s the unified frontend that makes having multiple agents actually manageable instead of chaotic. If you use two or more local AI coding agents regularly, this extension will save you real friction every day. At 512 stars and climbing, I’m not the only one who thinks so.