One agent just silently deleted your auth config. Another rewrote 47 files. aimux shows you exactly what happened, and lets you judge if it was good.
Quick Start
brew install zanetworker/aimux/aimux
aimux # terminal dashboard
aimux web # browser dashboard at localhost:9090
Requires tmux. Auto-discovers running Claude and Codex agents.
What It Does
See everything. Live trace of every prompt, response, and tool call as agents work. Watch file edits, bash commands, and API calls turn by turn.
Judge quality. Label turns GOOD, BAD, or WASTE. Add notes. Build eval datasets from real production sessions, not synthetic benchmarks.
Export anywhere. OTEL to MLflow or Jaeger for evaluation pipelines. JSONL to disk for offline analysis. Your annotations become feedback assessments.
Zero setup. No config, no instrumentation, no SDK. Run aimux and it finds every running agent on your machine.
Features
Discovery & Monitoring
- Auto-discovers Claude and Codex agents
- Live status: active, idle, waiting, error
- Per-agent CPU%, memory, token usage, and cost tracking
- Model-aware pricing (Opus, Sonnet, Haiku, GPT)
- Cross-session search ("which agent edited auth.go?")
Tracing & Code Review
- Turn-by-turn trace of prompts, responses, tool calls
- PR-style diff review panel with file tree hierarchy
- Inline diffs for Edit/Write tool calls with syntax coloring
- Live diff pane alongside session terminal (auto-refreshes)
- Annotate turns as GOOD / BAD / WASTE with notes
- Export to MLflow, Jaeger, or any OTLP backend
Two Interfaces
- TUI: keyboard-driven, split-pane trace + live session
- Web: card grid, filtering, session history, trace + diff panels
- Dark and light themes (web dashboard)
- Plugin system for custom dashboard tabs
Agent Management
- Launch and spawn agents from CLI or UI
- Star/pin sessions to find them later across restarts
- Profiles: save named flag bundles for repeated use
- Session history with browse, search, path filter, and resume
- LLM-powered title generation for untitled sessions
- Kubernetes support: run agents on K8s pods
Agent-Ready CLI
aimux is built for AI agents. Commands support --json for structured output (the mcp proxy commands and remote serve print text, and chat --json streams JSON lines), errors include valid values, and agent-context dumps the full CLI contract.
aimux agents --json # discover running agents
aimux sessions --list --json --limit 5 # search past sessions
aimux spawn claude --dir ./proj --wait --json # start and wait for completion
aimux agent-context # full CLI schema for agents
aimux profile save work --provider claude # save reusable config
aimux feedback "the X flag doesn't work" # report friction
aimux chat <session-id> "next step?" # continue an idle Claude session
aimux mcp proxy serve # lazy, shared MCP servers for all sessions
aimux remote serve # your sessions from other devices (API, phone web client at /app/, approval notifications), via tailscale serve
See Agent Usage for the full guide and SKILL.md for agent workflow recipes. The remote API's contract, for building clients, is api/remote-v1.openapi.yaml (each server also serves it at /v1/openapi.yaml).
Documentation
Full documentation at zanetworker.github.io/aimux
- Getting Started - install, first run, CLI commands
- Configuration - full config.yaml reference
- Agent Usage - structured output, profiles, feedback
- Web Dashboard - cards, filters, sessions, diff review
- TUI Keybindings - keyboard shortcuts
- Tracing & Annotations - trace view, OTEL, JSONL
- Launch Modes - tmux, direct, task-driven
- Adding a Provider - extend with new agents
- Kubernetes - run agents on K8s
Built With
Bubble Tea |
Lip Gloss |
charmbracelet/x/vt |
creack/pty |
OpenTelemetry |
Starlight
License
MIT
