Mission control for AI coding agents: see, steer and evaluate every session.
Docs · Install · Quickstart · Changelog
Four agents in one list, then one agent's trace beside its live session, a turn labeled GOOD, and cost per project.
You run five agents in five tmux panes and lose track of which one rewrote auth.go. aimux finds every running and past session, shows each one turn by turn (prompts, tool calls, diffs), and lets you label turns so the good and bad runs become eval data.
brew install zanetworker/aimux/aimuxRequires tmux. No config, SDK or instrumentation: aimux reads the session files Claude Code and Codex already write.
aimux agentsPID PROVIDER STATUS PROJECT MODEL
7 claude Active payments-api opus
25 claude Idle docs opus
13 claude Idle web opus
19 claude Waiting infra opus
Then open the dashboard:
aimux # terminal UI
aimux web # browser UI at localhost:3000- See every agent at once. Running and past Claude Code and Codex sessions in one list, with live status, CPU, tokens and cost, and search across all of them ("which agent edited auth.go?").
- Know what each agent actually did. A turn-by-turn trace of prompts, tool calls and diffs, with a PR-style review panel. Tracing →
- Turn real runs into evals. Label turns GOOD, BAD or WASTE and export them as MLflow feedback, to any OTLP backend, or as JSONL. MLflow →
- Run and reach agents anywhere. Launch them locally, in OpenShell sandboxes or on Kubernetes. From your phone, follow your Claude sessions, continue idle ones and answer approvals. Remote sandboxes →
| Area | What you get | Docs |
|---|---|---|
| Discover | Auto-discovery of Claude Code and Codex, live status (working, idle, needs you, error; the same words on the phone), cross-session search, starred sessions, generated titles | Web dashboard |
| Trace and review | Turn-by-turn trace, inline diffs, PR-style review panel, live diff pane beside the session terminal | Tracing |
| Cost | Per-agent and per-turn cost with model-aware pricing | Cost tracking |
| Evaluate | GOOD / BAD / WASTE labels with notes; export to MLflow, Jaeger or any OTLP backend, or JSONL | MLflow |
| Launch | Six-axis launcher, saved profiles, quick-launch directories, launches from Google Tasks, continue an idle session with aimux chat |
Launch modes, Tasks, Chat |
| Run remotely | OpenShell sandboxes with terminal, traces and kill; agents on Kubernetes pods | Remote sandboxes, Kubernetes |
| Reach | aimux remote serve: follow and continue Claude sessions from the iPhone app or the phone web client over tailscale, answer approvals, type into sessions open in iTerm2, and open a phone session on the Mac |
Remote access |
| Share MCP | aimux mcp proxy install: your stdio MCP servers run once in a shared daemon, each at its own endpoint under its original name (permission rules unchanged), started on first call, shared ones stopped when idle and browser ones kept for as long as their Claude Code session runs; aimux mcp serve: session search and resume tools for agents |
MCP proxy |
| Interfaces | Keyboard TUI with split panes; web dashboard with cards, filters and dark/light themes; desktop and terminal-bell notifications | TUI keys, Notifications |
| Extend | Plugin tabs, new providers, --json on every listing command, aimux agent-context for the full CLI contract |
Plugins, Agent usage, Add a provider |
Use the Claude Code or Codex built-in history instead if you run one agent at a time; aimux pays off when several run in parallel.
What it doesn't do: it doesn't plan or orchestrate multi-agent work, and it only supports Claude Code and Codex today (add a provider).
The web dashboard (aimux web): every agent by project, and one agent's trace.
| Task | Command |
|---|---|
| Search past sessions in all projects | aimux sessions "auth.go" --all --list --json --limit 5 |
| Start an agent and wait for it | aimux spawn claude --dir ./proj --wait --json |
| Resume a past session | aimux resume <session-id> |
| Jump to a running agent's iTerm2 pane | aimux jump <session-id or pid> |
| Continue a Claude session that is not open in a terminal | aimux chat <session-id> "next step?" |
| Export a trace | aimux export <session-id> |
| Follow Claude sessions from your phone | aimux remote serve (via tailscale serve) |
| Share MCP servers across sessions | aimux mcp proxy install --dry-run, then aimux mcp proxy install |
All commands
agent-context Machine-readable CLI surface for agents
agents List running AI agents
chat Send a message to an existing Claude session and stream the reply
collect Run headless OTEL collector
configs List named agent configurations
environments List configured environments
export Export a session trace
feedback Send feedback about the CLI
jump Bring a running agent's terminal pane to the front
kill Kill a running agent
mcp MCP server for session search and remote agents
profile Manage saved configuration profiles
remote API for your sessions from other devices, through tailscale serve
resume Resume a past session
sessions Browse and search past sessions
spawn Start a new AI agent session
version Show aimux version
web Launch web dashboard
Full docs: getting started, configuration.
Issues and PRs welcome. Run aimux feedback "..." to report friction straight from the CLI.