Agent multiplexer · a binary, not an app

Every agent. One calm terminal.

Keep Codex, Claude Code, OpenCode, and their shells running when your terminal disappears. Reattach over SSH, see blocked, working, and done at a glance, then jump straight to the pane that needs you.

Build from source Read the docs

Early access · source only · no Electron · no account · GitHub

The real Nagi agent cockpit filtering blocked, working, and done agents — Real Nagi build. Deterministic demo data. No mock interface.

opencodenagi+

 ▐▛███▜▌
▝▜█████▛▘
  ▘▘ ▝▝

Claude Code v2.1.198

Fable 5 with high effort · Claude Max

~/Projects/nagi/website

❯ make the Nagi hero mock look exactly like real claude code

● Ha. I read the real pane over the socket — the mock was on v2.1.168, Claude Code is v2.1.198 now. Missing the logo block and the whole rate-limit row.

● Plan:

· index.html — swap ✻ for the real ▐▛███▜▌ mark, bump the version

· css/style.css — coral bypass line, add the 5h / Week bars

❯ perfect. and make the dialogue about this, obviously

⠋ Baking… (13m 36s · esc to interrupt)

❯

~/Projects/nagi/website > main * ↑1 > ctx ──────── 3% 31k/1M

⏵⏵ bypass permissions on (shift+tab to cycle)

~/Projects/nagi main

❯ bun run dev

$ node scripts/prepare-docs.mjs && astro dev

02:10:44 [types] Generated 0ms

02:10:44 [content] Synced content

astro v5.18.1 ready in 668 ms

┃ Local http://localhost:4321/

Network use --host to expose

02:10:44 watching for file changes...

+ Thought: 468ms

→ Read src/main.rs

+ Thought: 292ms

This is src/main.rs — the entry point for nagi. Here's what it does:

1. Arg parsing — session::configure_from_args, then remote args for the SSH bridge.

2. Subcommand dispatch — server, client, update — each returns early.

3. Session mode — auto_detect_launch() connects to a running server or spawns one.

Build · MiMo V2.5 Free · 16.6s

Build · MiMo V2.5 Free OpenCode Zen

25.2K (13%) ctrl+p commands

migratelogs+

 ▐▛███▜▌
▝▜█████▛▘
  ▘▘ ▝▝

Claude Code v2.1.198

Fable 5 with high effort · Claude Max

~/Projects/llm-proxy

❯ prep the request-log migration and run it

● Wrote migrations/0007_request_log.sql — request_log table, retention index on created_at, daily rollup view.

● Local checks pass: sqlx prepare is clean, backfill dry-run touches 41,209 rows. Ready to apply.

Bash command

cargo sqlx migrate run

Apply pending database migrations

Do you want to proceed?

❯ 1. Yes

2. No, and tell Claude what to do differently (esc)

~/Projects/llm-proxy main

❯ tail -f logs/proxy.log

02:30:58 INFO request id=4818 route=/v1/messages 200 312ms

02:31:04 INFO request id=4819 route=/v1/messages 200 287ms

02:31:09 DEBUG cache hit ratio 0.93

02:31:12 INFO request id=4820 route=/v1/messages 200 301ms

02:31:15 WARN upstream latency p99 1.8s

02:31:18 INFO request id=4821 route=/v1/messages 200 295ms

02:31:21 WARN 429 from upstream, retrying in 2s

02:31:24 INFO request id=4822 route=/v1/messages 200 1102ms

02:31:30 DEBUG cache hit ratio 0.94

02:31:33 INFO request id=4823 route=/v1/embeddings 200 88ms

codexserver+

• Explored

└ Read live_handoff.rs

• Ran

└ cargo nextest run -E 'test(live_handoff)'

└ 14 passed · 0 failed · 41.3s

• The handoff suite passes. Pane PTYs survive server replacement, and long-lived workloads — including the http.server in the next tab — keep responding afterward.

• Next I can wire the soak test into nightly CI and fail on p95 regressions. Want that?

›

gpt-5.5 medium · ~/qmp · main · Context 6% used

~/qmp main

❯ python3 -m http.server 8080

Serving HTTP on 0.0.0.0 port 8080 ...

02:28:41 "GET /runs/latest.json" 200

02:29:03 "GET /runs/latest.json" 200

02:29:25 "GET /qmp/report.html" 200

02:30:07 "GET /runs/latest.json" 200

02:30:18 "GET /favicon.ico" 404

02:30:42 "GET /qmp/report.html" 200

02:31:09 "GET /runs/latest.json" 200

Intent, agents, and live terminals in one calm surface. Click the sidebar: it is a real layout.

# why nagi

Run agents where the work is. Attach from anywhere.

Agents run wherever the work is: a server, a Mac Mini, a sandbox VM, anywhere you can ssh. You attach from any terminal, even a phone. Desktop agent managers can't leave the machine with the GUI. Nagi can. Why switch from an app →

local

Your machine

Split panes, create tabs, and keep agents running while your terminal comes and goes.

$ nagi

over ssh

Like tmux, on the server

SSH in and run Nagi on the remote shell. The session stays after you detach, ideal from a phone SSH client.

$ ssh you@server
$ nagi

thin client

Native remote attach

A local client to a remote session. Installs Nagi on the host for you, bridges your local clipboard (including image paste), and keeps your keybindings.

$ nagi --remote workbox

# responsive tui

Mobile-first when the terminal gets small.

Nagi stays usable over SSH from a phone or tablet. The terminal view remains real, while narrow screens get a switcher built for touch-sized decisions.

Nagi agent session over SSH on a phone — agent session over SSH

Nagi responsive switch menu on a phone — responsive switch menu

screenshots taken with moshi ❤️ on iPhone

# why different

Most agent managers are apps. Nagi is a multiplexer.

tmux and Zellij own persistent terminal sessions but don't understand agents. Desktop agent apps understand agents but live on one machine. Nagi keeps the multiplexer in your terminal and makes it agent-aware. See the full comparison →

not a terminal emulator

Ghostty, Kitty, iTerm, Alacritty: your terminal stays.

not a browser dashboard

No web view, no account, no hosted control plane.

more than tmux for agents

Persistence plus clickable panes, agent state, an API.

Capability	tmux / Zellij	agent apps	worktree orchestrators	Nagi
Runs inside your terminal	✓	—	—	✓
Persistent PTY sessions	✓	limited	embedded	✓
Remote SSH attach	✓	limited	remote projects	✓
Semantic agent state	—	partial	workspace status	✓
Direct agent attach	—	—	—	✓
Agents can orchestrate it	scriptable	partial	workflow-owned	✓

Runs inside your terminal

tmux/Zellij ✓ agent apps — worktree apps — Nagi ✓

Persistent PTY sessions

tmux/Zellij ✓ agent apps limited worktree apps embedded Nagi ✓

Remote SSH attach

tmux/Zellij ✓ agent apps limited worktree apps remote projects Nagi ✓

Semantic agent state

tmux/Zellij — agent apps partial worktree apps workspace status Nagi ✓

Direct agent attach

tmux/Zellij — agent apps — worktree apps — Nagi ✓

Agents can orchestrate it

tmux/Zellij scriptable agent apps partial worktree apps workflow-owned Nagi ✓

# what you get

Real panes, agent state, and an API to drive them.

No rebuilt chat view. Real processes in real PTYs, with clickable layout, persistent sessions, and a control surface your agents can use themselves.

Real panes

Mouse-first panes, tabs, and workspaces. Keep your shell, fonts, and keybinds.

Agent state

Blocked, working, done, and idle at a glance across the whole session.

Persistence

Detach and reattach. Sessions and agents survive the terminal closing.

Control surface

A CLI and JSON socket API expose workspaces, panes, output, and waits.

# create workspace structure
nagi workspace create --cwd ~/project --label api
nagi tab create --label logs

# split a pane and run work
nagi pane split 1-1 --direction right
nagi pane run 1-2 "just test"

# wait, inspect, and continue
nagi wait agent-status 1-1 --status done
nagi pane read 1-2 --source recent-unwrapped

CLI reference → Socket API → Agent skill →

# build it

Run two agents. Close the terminal. Come back.

Nagi is early access and ships from source today. The terminal workspace and cockpit work now. The deeper mission loop is still being finished in public.

Build from source Current status