Using the terminal app

Using the terminal app (ts5-tui)

ts5-tui is TokenStation5 in a terminal: a colorful, keyboard-driven app for when you SSH into the PS5, or sit at its text console. It chats with the loaded model and keeps every chat, downloads, switches, checks and tunes models, and shows the PS5's GPU memory, temperatures and fan live. Linux edition only.

It is a client of TokenStation's own API (port 8081), like any other app. It loads nothing itself and uses no GPU memory, so it can run beside the TV app or the API-only service.

Availability: ts5-tui ships in the Linux packages from the linux-latest build of 2026-11-10 on: /usr/bin/ts5-tui with the .deb, .rpm and Arch packages, and ~/.local/bin/ts5-tui with the portable build. Already installed an older build? Run the install command again to update.

Start it

From your computer, SSH into the PS5, then run ts5-tui:

ssh ps5@<your PS5's address>
ts5-tui

The TokenStation title sweeps in for about a second (any key skips it), and you land in Chat.

Command	What it does
ts5-tui	Talks to the server on this machine (http://127.0.0.1:8081).
ts5-tui --url http://ps5.local:8081	Talks to another TokenStation (ps5.local alone works too: http and port 8081 are assumed).
ts5-tui --key KEY	An API key, when the server asks for one (the TV app's Require a key).
ts5-tui --theme Neon	Colors: Ocean (default), Neon, Sunset, Forest, Mono.
ts5-tui --no-intro	Skips the opening title.
ts5-tui --tmux / --no-tmux	Forces its own tmux session on or off (see tmux).

TS5_URL and TS5_API_KEY in the environment do the same as --url and --key.

Run it on the PS5 (over SSH) for the whole experience. Pointed at a PS5 from another machine with --url, chat works, but the sensors, models and logs it shows are that machine's, and loading a model is refused: models are switched where they are served.

The screen

 ◆ TokenStation5   F1 Chat  F2 Models  F3 Status  F4 Settings  F5 Logs        09:52   ← pages
 ● Qwen3-8B-Q4_K_M · helper Qwen3-0.6B · ctx 909/32k · 58.9 tok/s · GPU 5.7/13.6 GiB · die 80°C   ← live bar

   (chat, or the page you picked)

 Qwen3-8B-Q4_K_M  ctx 909 / 32,768
╭────────────────────────────────────────────────────────────────────────╮
│ ts5> your message                                                      │   ← the prompt
╰────────────────────────────────────────────────────────────────────────╯
 Enter send  Ctrl+J new line  Ctrl+N new chat  …                             ← keys for this page

The top bar stays put whatever you do below it:

Item	Meaning	Colors
● model	What the server has loaded.	Green: ready. Yellow: no model yet. Red: the server is offline.
helper	The helper (draft) model ts5-tune chose for it, or off.
ctx	Context used by this chat, of the model's total.
tok/s	The last answer's speed: any app's when ts5-tui can read the server's log, else its own chats'.
GPU	GPU memory used / total, in GiB.	Yellow under 2 GiB free, red under 1 GiB (the engine keeps 1 GiB back).
RAM	System memory used / total.	Same as GPU.
die	The chip's temperature (Tctl), the one the PS5's limiter watches.	Yellow from 78 °C, red from 83 °C.
fan	Fan speed.

On a narrow window the least needed items drop off first (fan, then RAM). When the die reaches 85 °C, where the PS5 starts slowing its GPU, a red warning line flashes under the bar: /fan max helps.

Five pages fill the middle. Switch with F1 to F5, Tab (next) and Shift+Tab (previous), or Alt+1 to Alt+5. On every page but Chat, 1 to 5 switch too, q quits, and / jumps to the prompt with a command started. From the prompt, /chat, /models, /status, /settings and /logs go to a page.

On a Mac, macOS keeps the F keys for itself (F3 is Mission Control). Use Tab / Shift+Tab, 1 to 5 on the pages, the commands above, or fn+F1 to fn+F5. Option+1 to Option+5 work when the terminal sends Option as Meta: in Terminal, Settings → Profiles → Keyboard → Use Option as Meta key; in iTerm2, Profiles → Keys → Left Option key: Esc+.

Chat (F1)

Type at ts5> and press Enter. The answer streams in with its formatting: headings, lists, quotes, and code blocks with highlighting. A thinking model's reasoning is folded away: while it thinks you see a pulsing thinking… line with its last words, then ▸ Thought for 4.1 s. Ctrl+T unfolds every chat's thinking (and folds it again). Under each answer: its speed, tokens, prompt speed and time.

Key	What it does
Enter	Send.
Ctrl+J or Alt+Enter	A new line in the message. Pasting several lines works too.
↑ / ↓	The messages and commands you typed before.
Esc	Stops an answer (Ctrl+C too). With no answer running, clears what you typed.
Ctrl+N	A new chat. The one you leave is kept (see Saved chats).
Ctrl+T	Show or hide the thinking.
PgUp / PgDn	Scroll the chat a page. Ctrl+↑ / Ctrl+↓ scroll a line, and so does the mouse wheel in tmux. End (with nothing typed) goes back to the bottom.
← → Home End	Move in what you're typing. Ctrl+← / Ctrl+→ by word.
Ctrl+A / Ctrl+E	Start / end of the line.
Ctrl+U / Ctrl+K / Ctrl+W	Delete to the start / to the end / the word before.
Ctrl+Q	Quit.

The top bar always shows how full the chat is (chat 7.8k/32k), and the line over the prompt reads "chat N of M tokens": yellow from 75%, red from 90%. When an answer stops because the chat filled the context, a note under it says so, with the numbers, and that Ctrl+N starts a new chat (the full one stays saved). For a model loaded at a fixed context, it adds that Settings → Context → auto gives it more. If the answer hit your Answer length cap instead, the note says that. A message sent to a chat that's already full gets "This chat is full" in place of an answer, with the same Ctrl+N tip.

Saved chats

Every chat is saved after each answer, the way the TV app saves them: one file per chat in ~/.local/share/tokenstation/chats/ (or $TS5_DATA_DIR/chats). Nothing to do: quit, come back, and they're there.

• /chats lists them, newest first, with when and how many messages. The TV service's chats (/var/lib/tokenstation/chats) are listed too when your user can read them.
• /chats TEXT lists only the chats that have every word of TEXT.
• /open N opens number N from the list, and you go on where it ended. /open TEXT opens the newest chat that has TEXT. Going on with a TV chat makes a copy of your own.

Commands

Type / and the matching commands show above the prompt. Tab completes the command, the model name after /model, /tune, /check or /bench, and the catalog name after /download (press Tab again to step through the matches).

Command	What it does
/model NAME	Loads another model (see Models). Part of the name is enough. /model alone opens the Models page.
/models	The Models page. /chat, /status, /settings and /logs go to the others.
/download NAME	Downloads one of the models tested on the PS5 (see Downloading models): qwen3-14b, qwen35-9b, deepseek-r1-14b, qwen3-8b, llama31-8b, qwen25-coder-7b, qwen25-coder-14b, gemma3-12b, gemma4-12b, llama32-3b, and the helpers qwen3-0.6b, llama32-1b, gemma3-1b. Tab completes the name. /download alone opens the list on the Models page.
/download URL [FILE] [--sha256 HEX]	Downloads any other .gguf. A ts5 fetch URL FILE --sha256 HEX line, as on Models tested on PS5, works pasted at the prompt as it is.
/new	A new chat (/clear too). The one you leave is kept.
/chats [TEXT]	Your saved chats, or the ones with TEXT (see Saved chats).
/open N	Opens saved chat N from /chats (or /open TEXT: the newest with TEXT).
/save [FILE]	Saves this chat as Markdown, thinking included. Without a file: ~/tokenstation-chat-YYYYMMDD-HHMM.md.
/find TEXT	Scrolls to the newest message containing TEXT.
/tune [NAME]	Runs ts5-tune on the loaded model (or NAME), live on the Models page.
/check [NAME]	The GPU check on the loaded model (or NAME): ts5 check, the GPU's answers against the CPU's.
/bench [NAME]	The speed test on the loaded model (or NAME): ts5 bench, prompt and answer tok/s.
/usage	The API's use over the last day and week, by app (its API key's name, else its address) and by model: requests, prompt and generated tokens, tok/s. From the TV or desktop app's usage.tsv.
/keys	The API keys: name, on or off, when made, when last used. Keys are made on the TV app's Settings page, and only their SHA-256 is kept, so they can't be shown again.
/think on, off, auto	Thinking before answering, for models that can switch it. auto leaves it to the model.
/fan auto, cool, early, max	The PS5's fan (see PS5 hardware on Linux). early ramps from 40 °C and is at full speed by 55 °C.
/theme NAME	Colors: Ocean, Neon, Sunset, Forest, Mono. Without a name, the next one.
/status, /settings, /logs	Those pages.
/help	Every command and key, in the chat.
/quit	Leaves (/exit and /q too, and plain exit, quit or q). The server keeps running.

Models (F2)

Every .gguf on this machine, biggest first, with its size, the helper ts5-tune picked for it, and the date it was tuned. ● marks the one loaded. Under them, ⇣ Download: the tested models you don't have yet. Below the list, the selected model's tuned settings, each with what it costs: 16- or 32-bit prompt math, the helper, the chat memory (automatic: full precision, or 8-bit where that fits a longer chat) and, for the loaded one, its context.

Key	What it does
↑ ↓, PgUp PgDn	Choose.
Enter	Load it. On a catalog row: download it.
T	Tune it with ts5-tune --models (10 to 20 minutes).
C	The GPU check (ts5 check): the model's answers on the GPU against the CPU's.
B	The speed test (ts5 bench): a 512-token prompt and 128 tokens of answer.
D	On a catalog row: download it.
X	Stop the download (it goes on from there next time).
R	Look for model files again.

It looks in /var/lib/tokenstation/models (or the service's MODELS_DIR), ~/.local/share/tokenstation/models (the desktop app's), ~/models, and $TS5_MODELS if set. Files under 1.5 GiB are listed as helper size and aren't tuned.

What Enter does depends on what serves the API:

Running	Enter
The API-only service (tokenstation)	Switches its model and restarts it, through the package's hardware helper. Your user must be in the tokenstation group (sudo adduser $USER tokenstation, then log in again). The model must be in the service's models folder.
Nothing	Starts ts5 serve with that model on 127.0.0.1, with the context set in Settings (auto by default: the longest the model fits, up to 32,768 for Qwen3 14B on the PS5). It keeps serving after you quit ts5-tui, which then prints its process number; kill it to stop it. Its log is ~/.cache/tokenstation/tui-serve.log.
The TV app (tokenstation-tv)	Nothing: pick the model on the TV's Models page.
Another program on port 8081	Nothing: switch the model there.

The top bar shows loading… with the seconds until the server reports the new model.

Helpers aren't loaded by hand. A helper is a small model that drafts tokens for a big one (same answers, faster, mostly on code), and it comes along with the main model as that model's tuning picked: T tunes a model and picks the fastest helper among the files here. The top bar shows the loaded model's helper (none until it's tuned), and the Models list shows each model's. Enter on a helper-size file (under 1.5 GiB) asks first, because it would replace the main model.

Downloading models

The list under ⇣ Download is every model on Models tested on PS5, best first, with the helpers last. Each is pinned to the exact upload that was tested, with its size and SHA-256. Models already on this machine aren't listed. Choose one and press Enter (or /download NAME from the prompt). After a helper downloads, press T on its big model: tuning picks the helper up. The row shows a bar and the speed; the box below shows where the file goes, how much space is free there, what it's doing (connecting, downloading with the time so far and left, checking it's the exact file), and its steps ticked off. The top bar shows the download's percent and time left on every page.

• It downloads with ts5 fetch into a .part file and keeps the model only when its size and SHA-256 match, so a broken download never shows up as a model.
• It checks the free space first and won't start when the drive can't hold the file.
• X stops it. The .part stays, and downloading it again goes on from there. Quitting ts5-tui stops it the same way.
• It goes to the service's models folder when your user may write there, else to ~/.local/share/tokenstation/models (or $TS5_MODELS). The API-only service only reads its own folder, so ts5-tui then shows the sudo mv command that moves the file there.

Any other model: /download URL [FILE] [--sha256 HEX], or paste a ts5 fetch line from Models tested on PS5 at the prompt as it is:

ts5> ts5 fetch https://huggingface.co/…/resolve/…/Model.Q4_K_M.gguf Model.Q4_K_M.gguf --sha256 5bc6f9…

It downloads the same way and shows under the catalog while it runs. FILE is a name only (it goes where downloads go); without it, the name at the end of the URL. Hugging Face page links (/blob/) work too. Without --sha256 only the size is checked.

Tuning, the GPU check and the speed test

Tuning (T), the GPU check (C) and the speed test (B) open a progress panel:

• Working: what it's doing right now, in plain words, with a spinner, the test answer it's on, the time so far and about how long is left (the estimate shows after the first test answer).
• A checklist of every step it will go through, ticked off as each one finishes, with its result and how long it took: for tuning, checking the model gives correct answers, measuring its speed on its own, trying each helper model, saving the best settings. Skipped steps say why (for example, a helper whose vocabulary doesn't match).
• At the end, what it found: for tuning, the helper it picked and the speed before and after ("It writes 46% faster with the helper …: 58.9 tok/s, up from 40.3 on its own, with the same answers."); for the check, whether the GPU's answers match the CPU's; for the speed test, how fast it reads prompts and writes. The same line goes into the chat.
• Under it, the tool's own lines, for a bug report.

While a job runs, the top bar says so in yellow (tuning MODEL · step 3 of 7 · 2:13) instead of offline, on every page. Esc hides the panel and the job goes on (T shows it again), X stops it (nothing is changed then), PgUp/PgDn scroll. The GPU check runs the CPU side too, which takes a few minutes the first time for a model and is saved after. They need the GPU's memory: a ts5 serve that ts5-tui started is stopped for them. If the service holds a model, stop it first (sudo systemctl stop tokenstation). ts5-tune saves the results in your own ~/.config/tokenstation/tune/. The API-only service runs as its own user and reads /etc/tokenstation/tune/, so copy the file there for the service (sudo mkdir -p /etc/tokenstation/tune && sudo cp ~/.config/tokenstation/tune/<model>.gguf.env /etc/tokenstation/tune/).

Status (F3)

btop-style bars for this machine: GPU memory (with GiB free), RAM, the die (Tctl), board and GPU edge temperatures, fan (percent, rpm and mode), GPU load and clock. Below them, the die's temperature over the last minutes, and the generation and prompt speeds of recent answers: every app's when ts5-tui can read the server's log (see Logs), else its own.

F steps the fan through auto, cool, early and max (as /fan does).

Settings (F4)

↑ ↓ choose, ← → change, Enter edits a text setting (Esc cancels). Every change is saved at once, in ~/.config/tokenstation/tui.conf.

Setting	What it does
Server	The TokenStation to talk to.
API key	Only when the server asks for one. Kept in the settings file, readable by you only.
Theme	Ocean, Neon, Sunset, Forest, Mono.
Thinking	The model's choice, on, or off.
Temperature	The model's own, or 0.1 to 1.5. Lower is steadier.
Answer length	No limit, or a cap of 512 to 8,192 tokens. Under 1,024 turns thinking off.
Context (own)	The context for a model ts5-tui starts itself: auto (the default: the longest that fits), or 4,096 to 32,768. A setting saved as 8,192 by an older version reads as auto.
Fan (PS5)	auto (quiet), cool (keeps long prompts under the 85 °C limiter), early (ramps from 40 °C, for long prompts back to back), max (loud).
tmux	Its own tmux session from a plain terminal.
Opening title	The title at start.

The engine's own switches (prompt math, cache, helper) are per model, and ts5-tune sets them: Models → T.

Logs (F5)

What the engine writes, colored: errors red, warnings yellow, speed lines green. PgUp/PgDn and ↑ ↓ scroll, End follows the newest again, Home goes to the start.

It follows the API-only service's journal (journalctl -u tokenstation; reading it may need the systemd-journal or adm group), the TV app's /var/lib/tokenstation/ts5.log, or the log of the ts5 serve it started, whichever is running. TS5_TUI_LOG=/path/to/file ts5-tui follows a file of your choice.

tmux: SSH that can drop

From a plain terminal (an SSH session from Terminal, iTerm, PuTTY…), ts5-tui starts itself inside its own tmux session, called ts5, when tmux is installed (the packages recommend it). If the connection drops, nothing stops: the answer finishes, the model keeps loading, tuning goes on. Run ts5-tui again and you're back where you were.

• Ctrl+A then D leaves it running in the background; ts5-tui comes back to it.
• The mouse wheel scrolls the chat. To select text with the mouse, hold Shift (Option on a Mac in iTerm).
• It uses its own tmux server and settings (~/.cache/tokenstation/tmux.conf): your ~/.tmux.conf and your other tmux sessions are left alone, and tmux's own status bar is hidden.
• Quitting ts5-tui ends the session.

In an IDE's terminal (VS Code, Cursor, JetBrains, including their Remote-SSH) tmux stays off, because the IDE keeps its terminals alive itself and its shortcuts would clash. --tmux turns it on there anyway, and --no-tmux or Settings → tmux turns it off everywhere.

Plain SSH leaves more memory for models. An IDE's Remote-SSH server runs on the PS5 and takes memory the GPU could use (Cursor's took 4.7 to 6.5 GB). For coding, run the IDE on your computer and point it at the PS5's API instead (Connecting apps).

When you quit

Ctrl+Q (always shown at the bottom right), exit, quit or q at the prompt, /quit, or q on a page other than Chat. In ts5's tmux session, Ctrl+A d leaves it running instead. The chat is printed into your terminal, so it stays in its scrollback (also after a tmux session). The server and its model keep running.

If something's off

What you see	Why, and what to do
● offline in the top bar	Nothing answers at the address. Start the service (sudo systemctl start tokenstation) or the TV app, or load a model on the Models page, which starts ts5 serve. Check the address in Settings.
needs an API key	The server has Require a key on: put a key in Settings → API key, or start with --key.
No temperatures, fan or GPU memory	ts5-tui isn't running on the PS5 (a Mac, or --url to another machine), or the PS5 kernel's sensors are missing. Run it over SSH on the PS5.
Load or fan: "needs the tokenstation group"	sudo adduser $USER tokenstation, then log out and in.
Colors look washed out or wrong	The terminal has no 24-bit color, so ts5-tui uses 256 colors. Terminals that support it set COLORTERM=truecolor; try another theme, or Mono.
Odd characters instead of boxes	The terminal isn't UTF-8. Set a UTF-8 locale (export LANG=C.UTF-8).
tmux opens when you don't want it	ts5-tui --no-tmux, or Settings → tmux → off.

More in Troubleshooting.