The real bottleneck with coding agents is coordination, not capability

The core problem Multica solves is coordination overhead, not capability. You already have Claude Code, Codex, Cursor — the README opens by naming exactly this situation: multiple agent CLIs running in separate terminal tabs, each one stateless, each one requiring you to re-explain context every session. The more agents you run, the more time you spend managing them rather than reviewing their output. Multica's answer is to treat agents as board members rather than command-line tools.

The mechanism is a local daemon that sits next to your code on whatever machine you designate as a runtime. The Go backend communicates with that daemon over WebSocket, dispatching tasks and streaming back execution logs. When you assign an issue to an agent, the daemon spawns the appropriate CLI — one of 23 supported, from Claude Code and Codex to Kimi, Grok, Qwen Code, and MiniMax Code — runs it, and attaches the full execution trace (every tool call, every command, every error, timestamped) back to the originating issue. The intent, the run, and the diff stay linked. Nothing ships until a human approves it at the review gate.

The 23-runtime count is the most immediately practical detail here. Multica doesn't ship a model or a provider; it drives whatever you've already installed and authenticated. Switching from one agent CLI to another is a dropdown selection, not a migration or a rewrite. That design choice also means the project stays neutral on which underlying model wins — a reasonable hedge given how fast that landscape moves.

The "Skills" feature deserves attention: a solved problem becomes a reusable playbook that any agent in the workspace can draw on. Combined with "Squads" — mixed teams of agents and humans where a leader routes work — this starts to look less like a task runner and more like a lightweight organizational layer for hybrid teams.

The architecture is straightforward: Next.js frontend, Go backend with Chi and gorilla/websocket, PostgreSQL 17, an Electron desktop app sharing the web UI packages, and an Expo/React Native iOS client that currently builds from source rather than shipping through the App Store. Self-hosting runs via Docker Compose or Helm. The license is Apache 2.0 with additional conditions around hosted services and commercial embedding — worth reading before building a product on top of it.

The name is a deliberate callback to Multics, the 1960s time-sharing OS that let multiple users share one machine as if each had it alone. The README frames agents as making time-sharing relevant again, this time with both humans and machines as the concurrent users. That framing is more than branding — it's the actual design thesis, and it's a coherent one.