The AI developer tooling space is moving fast enough that categories are still forming. Tools that look similar on the surface — "AI terminal things" — actually solve different problems at different layers. Understanding where gmux fits requires being precise about what layer each tool operates on.
L3 — the interaction layer — only becomes a real problem when you have enough agents that moment-to-moment management becomes the bottleneck. That threshold is higher than most people have hit yet, but it's coming.
Layer 1 · orchestration.
Multica (27.3k ⭐) — "your next 10 hires won't be human." A project management layer where AI agents are first-class assignees in task boards. Auto-detects OpenCode (and therefore qalcode2) on PATH, and Hermes, and Claude Code. Self-hostable.
Hermes Agent (145k ⭐) — self-improving CLI agent from Nous Research with its own messaging gateway, persistent memory, parallel subagents, cron scheduler. Also one of Multica's supported runtimes.
How they relate to gmux: Multica answers "what's queued for which agent." gmux runs underneath, answering "what does the human do right now with the agent that just got a task." They don't compete. The stack:
Multica → assigns task to qalcode2 agent
↓ agent runs in tmux window
gmux → shows state, handles gesture/voice/phone interactionLayer 2 · execution.
qalcode2 is the primary execution runtime in the gmux ecosystem — a fork of opencode with non-AVX hardware support and the HTTP API gmux reads. Every gmux feature that surfaces AI state is reading data qalcode2 generates.
opencode, Claude Code, Aider, DeepSeek-TUI — all supported. The integration for any agent with a similar HTTP API is straightforward; agents without one fall back to pattern matching.
How they relate to gmux: qalcode2 is excellent at single-agent work. gmux is what you add when you're running ten of them. The clean framing:
qalcode2 = single-agent executor.
gmux = multi-agent interaction layer.
Layer 3 · interaction. (gmux's lane.)
DeepSeek-TUI gained 21,752 GitHub stars in one week in May 2026 — the clearest signal yet that terminal-native AI tooling is hungry for innovation. But every viral entrant so far is single-agent. One model, one conversation, one pane.
The interaction problem at multi-agent scale is what gmux is built around. The three pillars — gestures, agent management, visual oversight — are all interaction-layer concerns. None of the other tools in this space touch them.
Gestures as a real input device.
Hands as a second input alongside the keyboard. MediaPipe local, two hands with separate roles, dwell-select, two-mode (passive while typing, active for navigation).
Agent management that keeps work flowing.
SSE-based state detection, six-state colour code, todo progress per window, permission roundtrip across the workspace, 600-token memory context injected into every new pane.
Visual oversight for the whole fleet.
Multi-agent grid, agent flowchart (attention paths), memory panel (decisions over time), phone companion. Four lenses on the same underlying data.
Nothing else in this space combines these three. That combination is the moat.
The honest summary table.
| Tool | Layer | Multi-agent | Gestures | Memory | Phone | Local |
|---|---|---|---|---|---|---|
| gmux | L3 · Interaction | ✅ core | ✅ MediaPipe | ✅ 3-layer | ✅ companion | ✅ |
| Multica | L1 · Orchestration | ✅ | ❌ | ❌ | ❌ | ✅ self-host |
| Hermes Agent | L1+L2 | ✅ subagents | ❌ | ✅ persistent | ❌ | ✅ |
| qalcode2 | L2 · Execution | ❌ single session | ❌ | ❌ | ❌ | ✅ |
| DeepSeek-TUI | L3 (single-agent) | ❌ | ❌ | ❌ | ❌ | ✅ |
| Warp | L3 (single-agent) | ❌ | ❌ | ❌ | ❌ | ❌ cloud |
| tmux | L0 · Infra | ✅ dumb | ❌ | ❌ | ❌ | ✅ |
The risk we're actually running.
It's not competition from similar tools. Nothing similar exists right now. The risk is that qalcode2 (or its successor) grows a multi-agent workspace mode natively, absorbing the status-bar and session-management features.
Gestures and voice are hard to absorb into a coding agent tool — they require camera/mic access and a separate visual layer. Same for the phone companion, the memory layer, and the projector mode. That's the durable differentiation.
The window to establish gmux as the standard multi-agent terminal layer is open and currently uncontested. DeepSeek-TUI's growth proves the audience exists. The Python stack already ships. The main blocker is publication.
Try the demos.
All four run in your browser on simulated data with no backend — the same UI code that ships in the desktop app, just disconnected from real data sources.
gmux: pip install gmux · gmux.ai