MARS

Mission control for your terminal — a non-modal, Emacs-compatible terminal editor with a Claude-Code-style mission-control command bar, a built-in LLM agent, real terminal panes, and tmux/zellij-style persistent sessions. One tool, one set of keys.

███╗   ███╗ █████╗ ██████╗ ███████╗
████╗ ████║██╔══██╗██╔══██╗██╔════╝
██╔████╔██║███████║██████╔╝███████╗
██║╚██╔╝██║██╔══██║██╔══██╗╚════██║
██║ ╚═╝ ██║██║  ██║██║  ██║███████║
╚═╝     ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝

Build & install

Linux/macOS:

source ~/.cargo/env            # if cargo isn't on your PATH
cargo build --release
# put it on your PATH:
ln -s "$PWD/target/release/mars" ~/.local/bin/mars   # or copy it anywhere
mars --selfcheck               # optional: run the built-in test suite

Windows PowerShell (Rust's MSVC toolchain and Visual Studio Build Tools required):

cargo build --release
.\target\release\mars.exe --selfcheck

Quick start

mars                    # start a session: the MARS banner, then a shell in your cwd
mars notes.md           # edit a file (also inside a session)
mars -s notes.md        # standalone mode: no session daemon, just edit
mars help               # full CLI reference

Sessions are the default, tmux-style: a bare mars gets an auto-numbered session that survives closing the window. Press any key to dismiss the startup banner.

Inside the editor, four keys carry you everywhere:

C-g cancels anything. Every menu row shows its real keybinding, so the fast path teaches itself as you go.

The workspace board ranks every workstream so "anything need me?" is answerable at a glance: blocked ⏸ and failed ✗ sort to the top, then running, done ✓, idle — each with one status bubble (● colored by state) and a plain-English status line. Select a row and press s for an on-demand, one-line "what is this doing?" summary (a single low-tier model call, guarded so it never over-fires). Any Markdown buffer can also toggle into a read-only reading-mode — real wrapping, tables, and nested lists, dressed in the Mars palette, scrolled with the editor's own motion keys (↑/↓, ⌥↑/⌥↓, M-</M->); toggle again to edit.

Moving around fast

Fast motion is bound to ⌥ (Option) and to C-x jumps that work on every terminal (enable "Use Option as Meta" if your terminal offers it).

Search doubles as teleport. C-s, type a word you can see — it jumps as you type, with a 3/12 counter. Press Tab and every visible match gets a one-letter label; press a label to teleport straight there. And you don't have to press Enter: just start editing or hit a motion key — the search commits at the current match and your key applies. C-s/C-r cycle matches, C-g cancels (restoring where you started).

Sessions — replace tmux/zellij

Sessions keep your buffers, panes, and running shells alive when the window closes, the SSH connection drops, or you just walk away.

mars new work           # start (or re-attach) a session named "work"
mars ls                 # what's running, and whether anything is attached
mars attach             # reattach the most recent session
mars attach work        # reattach a specific one
mars rename work api    # rename a running session (live — nothing restarts)
mars kill work          # end a session from outside (autosaves first)

The daily rhythm:

1. Start: mars or mars new work — everything from here on lives in the daemon.
2. Detach when you want the terminal back: press C-t then D — or just close the window. Both leave shells running and buffers intact.
3. Come back: mars attach (or mars attach work). Your layout, buffers, and that build you left running in a terminal pane are exactly where you left them. If anything happened while you were gone, the Mission Briefing boots up like a console coming online — a mission clock, a plain-English situation report in the mission-control voice that types itself in ("The trainer went down at epoch 3 — CUDA OOM, needs a smaller batch. The build came home green."), then a systems board of every workstream (failures first, then blocked ⏸, done ✓, running) with a "why" line under anything that failed and a teal ★ on a long run that finished clean. Any key resumes exactly where you left off. Each briefing is remembered, so the next return reports progress against it. Nothing happened → no briefing. (mission_briefing = 1 swaps it for a one-line notice; C-x g still opens the full Away Digest timeline anytime.)
4. Finish for real: quitting (C-x C-c) just detaches — a session only ends when you kill it: Kill session in the command bar (confirm-gated), mars kill work from outside, or mars killall to sweep every session and start fresh.

mars ls tells you the state at a glance:

SESSION              STATUS
work                 detached — reattach: mars attach work
review               attached

Safety nets, on by default: modified files autosave every 30s and on every detach/disconnect (scratch buffers are never touched), and each daemon logs to ~/.local/state/mars/<name>.log — if a session ever dies, the postmortem is there.

Notes: one client per session — attaching from a second window takes over from the first (it gets a clean "another client attached" message). Attaching from a different-sized terminal just reflows.

The agent

Beta. The AI features — the ? ask flow, agent-proposed RUN:/TYPE: directives, refactors, triage, watch summaries, and the away digest — are new and still being tuned. Treat the agent as an assistant, not an authority: review what it proposes before running it (destructive actions are gated, but read them). The core editor, multiplexer, and sessions are stable.

Works out of the box with a free-tier key from any of:

export ANTHROPIC_API_KEY=...   # Claude — defaults to claude-haiku-4-5
export OPENAI_API_KEY=...       # OpenAI — defaults to gpt-4o-mini
export GROQ_API_KEY=...         # Groq (free tier) — defaults to qwen/qwen3-32b
export GEMINI_API_KEY=...       # Google AI Studio (free tier) — gemini-3.1-flash-lite
# or any OpenAI-compatible endpoint (e.g. local Ollama):
export MARS_LLM_KEY=... MARS_LLM_URL=http://localhost:11434/v1 MARS_LLM_MODEL=llama3
# override the model for any provider:
export MARS_LLM_MODEL=qwen/qwen3-32b

Detection is paid-first: if several keys are set, Claude → OpenAI → Groq → Gemini (an explicit MARS_LLM_KEY always wins). Cheap defaults per provider — reach for a bigger model with MARS_LLM_MODEL, not by default.

Calibrating cost/latency (debug mode). To see where tokens and time actually go — so you can right-size the model per task — run with --llm-debug (or MARS_LLM_DEBUG=1) to log every call, then profile it:

mars --llm-debug              # logs prompts, models, tokens, latency to ~/.mars/logs/
mars llm-stats                # per task×model, ranked by total tokens:
#   TASK       MODEL            N  AVG_IN  AVG_OUT  TOT_TOK  %TOK  AVG_MS  ERR
#   watch      qwen/qwen3-32b   2    6050      195    12490   62%    3025    0   ← heaviest
#   ask        qwen/qwen3-32b   2    1720      215     3870   19%     900    0
mars llm-stats --raw          # full inputs/outputs per call
mars llm-stats --daily        # day-by-day token trend (a bar per day)
mars llm-stats --json         # machine-readable (rows + a daily series)
mars llm-stats --since 7d     # only the last 7 days (also 12h, 30m)

To keep it on without exporting the env var every time, set it in the global MARS config, ~/.mars/config.json (alongside the rest of Mars's state) — like a shell rc:

{ "env": { "MARS_LLM_DEBUG": "1" } }

Its env entries are exported at startup (the real environment still wins), so the session daemon inherits them too.

Reasoning models (Qwen3, DeepSeek-R1) work — their <think> blocks are stripped from answers automatically.

Then ? in mission control, or from the shell:

mars ask "how do I move a pane to the other side?"

The agent works on every box — your key never leaves home

Beta. The SSH features (mars ssh, mars keyd, the fleet view, and the remote installer) are new and still being hardened — as are the AI features they carry (see The agent). The core editor, multiplexer, and sessions are stable; the remote/tunnel path may have rough edges — please report anything you hit. Native Windows can be the home host when the remote is Unix. It uses stock OpenSSH without ControlMaster: a short bootstrap connection checks/installs Mars, then one foreground connection owns the session and tunnel. Password/2FA users may therefore authenticate twice. Windows as the remote host is pending.

You set your key once, on your own machine, and the agent works on every host you SSH into — without the key ever landing on a remote box (not in its env, not in its shell history, not on its disk).

mars ssh gpubox           # ssh in — forwards the auth socket AND auto-starts the
                          # key broker if needed (inheriting this shell's API key).
                          # `mars` on gpubox → the agent just works. No key on the box.

(The broker — mars keyd — starts on demand the first time you mars ssh; run it explicitly only if you want it in a specific shell.)

On a Windows home host, the bootstrap and interactive connections authenticate independently. The attached remote session receives the current tunnel route during its attach handshake, so detach and reattach preserve both the workspace and broker access.

Installing mars on a fresh host. Mars needs a modern Rust toolchain (≥ 1.85) — a distro-packaged cargo (e.g. Ubuntu's 1.75) is too old and fails with a cryptic edition2024 error. Don't apt install cargo; install rustup (the official way, from rust-lang.org/tools/install), then install mars from crates.io:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh   # Rust toolchain
. "$HOME/.cargo/env"
cargo install mars-terminal                                       # the `mars` binary

(install.sh in this repo automates exactly those steps, including detecting a too-old distro cargo. mars ssh stages this script and runs it when Mars is missing, or when a Windows-home handoff needs a newer protocol.) Your API key never lands on the box — it's served from home over the tunnel.

The remote never makes the LLM call itself — it proxies the request home through the SSH tunnel, and the completion comes back. Compromise the box and there's nothing to steal; close your laptop and remote access ends with the tunnel. Jump hosts, ProxyCommand, and hardware keys all work (it wraps your real ssh). mars ls shows the hosts you've been on, newest first — type a number or a name to hop back.

The agent sees your screen — editor buffers, terminal output, your layout — so "why did this build fail?" needs no copy-paste. It holds a conversation (C-l starts a fresh one), and it can act: RUN: fires an editor action, TYPE: types a shell command into your terminal pane, OPEN: jumps to a path:line from a stack trace — always shown first, always one explicit Enter away, never automatic.

What that unlocks:

• Ask about a selection. Select code (Shift+arrows or C-x h), then ? — the exact selection goes along as context, not just the visible screen.
• Reversible refactors. Ask "simplify this" / "add error handling" on a selection: the panel shows ▶ Enter to replace the selection (N lines). Enter applies it as one undo step, so a single C-/ reverts the whole AI edit.
• Triage. C-x ? (or "why did this fail?" in the bar) grounds the agent in the focused terminal's output; C-x e explains what's at the cursor.
• Watch a pane. Kick off a long command, then C-t w. When it exits or goes quiet (~20s), Mars leaves a one-line verdict at the bottom, failures first (✗ failed: linker error · build). Esc dismisses. This fires even while detached — the daemon keeps watching, so mars attach later lands on the verdict.
• Ask beyond the visible screen. Questions like "when did this first start failing?" or "does the error in the api tab match this code?" let the model request more context (the pane's full scrollback, or another tab); Mars supplies it and re-asks once, silently. You do nothing special.
• Shell translation. In a terminal pane, Ctrl+Space then plain English ("find big files here") → the agent translates it to a shell command and shows it for you to confirm. Typed a real Mars command instead? It's recognized and run directly. (! still forces shell, ? asks, @ opens Navigator.)

With an agent connected, tabs you haven't named get a quiet auto-generated label from their content (rename one yourself and it's yours forever; auto_name_secs = 0 turns it off).

Keys you already know

Mars speaks three dialects at once — whichever your fingers know:

• Navigator (browse files): C-x C-f — or Ctrl+Space then @ — opens Navigator, the file sidebar on the left. Folders are bold + colored and collapsed — arrow to one and Enter/→ expands it in place (← collapses); on a file, → previews it (reversible) and Enter opens it; ../ at the top steps up a directory, and Ctrl+Space on a folder re-roots into it (descend — the mirror of ../). Press . to show/hide dotfiles (the important things often live in hidden folders). Start typing to fuzzy-filter the whole project to a shortlist; Esc closes.
• Emacs: C-x C-s save · C-x C-f open · C-s isearch · M-% query-replace (y/n step, ! all) · C-k/C-y kill/yank · C-x 2/C-x 3/C-x o windows · M-x mission control
• Modern/Mac: C-c/C-v copy/paste (system clipboard) · Shift+arrows select · typing replaces selection · Tab/Shift-Tab indent/dedent a selected block · mouse click/scroll/wheel
• tmux/zellij: C-t space warp · M-{/M-} or C-PgUp/PgDn switch tabs · M-1..9 jump to tab · C-o/Ctrl+arrows move between panes · C-|/C-- splits · scrollback with the wheel or Shift+PgUp/PgDn

Undo, two ways. C-x u (or C-/) undoes — a typed run coalesces into one step, and an applied AI refactor is always exactly one step. M-/ redoes. For bigger jumps, C-u opens time-travel: ←/→ scrub back and forward through your edit history, Home rewinds to the very start, End returns to now, Esc done.

Everything is remappable in ~/.config/mars/keys.json; behavior knobs (autosave interval, scrollback depth, colors, timings, watch quiet threshold) live in ~/.config/mars/tuning.json, each with a plain-English description of what it does. Broke your config experimenting? mars reset restores default keys + tuning (your old files are kept as *.bak).

Color themes (beta). mars theme list shows the bundled themes — Mission Control (the default), Eclipse (high-contrast), Paper (warm light), and Hacker (green-on-black); mars theme <name> switches (recorded in ~/.mars/config.json). Or the Theme ▸ picker in the command bar switches live. Drop your own token→color JSON in ~/.mars/themes/. A running session keeps its look until it's restarted (or you switch from inside it). Every color is a single semantic token now, so a theme repaints the whole UI at once. A colored theme (Paper, Hacker) paints a solid background everywhere; the default honors your terminal's own background. Set opaque_background = 0 in tuning.json to keep the terminal background (transparency) even under a colored theme. It's new — please report anything that reads wrong (especially on the light theme).

Troubleshooting

• Staircase output (lines drifting right, like mars help printing diagonally): your shell's terminal was left in raw mode — usually by a force-killed program. Run any mars command (it repairs the terminal automatically on startup) or stty sane.
• M-… keys do nothing (macOS): enable "Use Option as Meta" in Terminal/iTerm — or use the Ctrl-based twins (C-o, Ctrl+arrows), which always work.
• A session shows dead (cleaned up) in mars ls: the daemon crashed or the machine rebooted. Check ~/.local/state/mars/<name>.log for the reason; autosaved file changes are already on disk.
• Fancy chords (C-{, C--) don't fire: they need a kitty-protocol terminal. The Alt-based twins work everywhere.

More

• architecture_overview.md — a file-by-file tour of the code: what lives where and how the pieces connect.
• DESIGN.md — architecture rationale, tradeoffs, and how the pieces fit.
• key_design.md — the design doctrine and product vision (why the keys are what they are, and where Mars is going).
• AGENTS.md — instructions for AI coding agents working on Mars.