wme.mov

wryme

(pronounced "write me", said quickly.)

wryme • that small, calm window where agents come to meet you!

You type your message at the top. The AI's reply appears right under what you typed and grows downward as it writes. Your earlier conversations sit below that, oldest at the bottom. No tabs, no menus, no surprises.

The little command you'll type to open it is wme.

Try it before doing anything

You don't need a key, a login, or an internet connection to see what this thing looks like. Just open it in your browser, or install it (below) and run wme. With nothing set up, it answers in canned nonsense so you can poke around the screen and decide whether you want more.

Installing it

cargo install wryme

or build yourself:

make build
make install

That puts a program called wme in your ~/.local/bin folder. If that folder isn't already on your PATH, your terminal won't find it; ask whoever set up your computer to add it, or add this line to your ~/.zshrc or ~/.bashrc:

export PATH="$HOME/.local/bin:$PATH"

The two pieces: shops and stations

wryme uses two simple ideas.

• A shop is a place that runs AI for you. It has an internet address and (usually) a key. The shop lists which models it can run. Examples: OpenAI, a server in your garage, or device at hand.

• A station is the AI you actually talk to. It picks a model from a shop and sets a few knobs (how creative, how careful, how long it can talk). You can save as many stations as you want and pick which one you want when you launch.

At launch wryme finds a shop that runs the model your station asked for, and away you go.

Set up your first shop

Edit the file ~/.config/wryme/shops.toml and put one block in it for the AI service you have an account with:

💡 Default protocol is open-responses, and chat-completions AKA OpenAI-Compat is also supported.

Open AI itself

export OPENAI_API_KEY="sk-...your-key..."

[[shop]]
name = "openai"
url = "https://api.openai.com/v1"
key_env = "OPENAI_API_KEY"

You can use key_env to point at any env var name, or key = "..." if you want to put the key directly in the file.

Grok

[[shop]]
name = "groq"
url = "https://api.groq.com/openai/v1"
key_env = "GROQ_API_KEY"
models = ["llama-3.3-70b-versatile", "llama-3.3-8b-instant"]

Ollama on your own computer. No key needed. Models auto-discover!

[[shop]]
name = "ollama"
url = "http://localhost:11434/v1"

Any other local runner, like antirez/ds4 (hint: this one runs better w/ chat-completions)

[[shop]]
name = "ds41"
url = "http://127.0.0.1:8000/v1"
protocol = "chat-completions"
models = ["deepseek-v4.1-flash"]

With just one shop and station configured, you can simply wme in, and ask the assistant to configure shops and stations.

💡 For agents: The models list works as a filter: if shop auto discovery doesn't work, this list is used; And if the discovery does work, it limits down to configured list. List them newest-first; if you launch wryme without picking a station, it will use the first model in the first shop.

Make your stations

Make the file ~/.config/wryme/stations.toml. Each station is a model plus optional dials:

[[station]]
name = "quick chat"
model = "gpt-4o-mini"

[[station]]
name = "deep thinker"
model = "o1-mini"
patience = "slow"

[[station]]
name = "creative"
model = "gpt-4o"
boldness = 1.3

The station dials, all optional:

• boldness: how wild the AI gets. A number between 0 and 2. Try 0.3 for safe, 0.7 for balanced, 1.3 for spicy. Unset means the AI decides.
• patience: how hard the AI thinks before answering. quick, steady, or slow. Only matters on models that support extended thinking. Unset means the AI decides.
• brainy: how much of that thinking gets shown. hush (show none at all), murmur, chatty, or gabby. Unset means the AI decides. A short retelling of the thinking prints under brain; the thinking itself, on models that hand it over, prints under heart.
• tinker_keep: how many tool pairs survive in replay. all (default), a positive integer (8, 3), 0 (keep none), or a percent (50%). Keeps the last N pairs, pair stays atomic.
• tinker_clip: how much of each kept tool result body survives. all (default, verbatim), a positive integer chars (500, 1000), 0 (empty body keeps call_id), or a percent (50% of original). Applies to replayed history, not the turn in progress.
• tinker_depth: how many rounds of tool use one turn may take before wryme forces an answer. all (default, unbounded), or a positive integer (8). At the limit the agent gets one more request with tools withheld and a nudge to write up what it has. Bounds latency and context growth; a tool-hungry model otherwise re-sends the whole conversation every round, getting slower each time with nothing printed.
• voice: speech voice for read-aloud assistant. Any say voice on macOS (Tara at 260wpm when unset), any espeak voice on Linux. Needs say/spd-say on PATH; silently off otherwise.

Usage preferences

These are saved in ~/.config/wryme/prefs.json. Change them from the preferences rows in the Ctrl-S popup, or edit the file:

{
  "accent": "cyan",
  "nametags": "inlined"
}

• accent: the highlight color. cyan (default), azure, green, lightgreen, yellow, magenta, red, blue, white, gray, lightcyan, lightyellow, or lightmagenta.
• nametags: how messages are labeled. standout (default) puts a you / assistant header above each message. inlined drops the header: a small icon sits on the left and the time on the right, with the text wrapped between them. ⏹ marks your messages. The assistant's icon shows what it's doing and blinks while busy: ⬩ writing, ✦ thinking, ‣ using a tool, ಄ reminiscing, ⌃ done. Thinking text is marked ✶ (brain) or ❤︎ (heart). A message you send while a reply is still coming in shows dimmed with waiting… until it goes out.

These reset every session:

view mode: Ctrl+T to toggle line-by-line vs pagination based scrolling.

voice toggle: Ctrl+V to toggle assistant voiceover whenever needed.

Using it

Type in terminal or desktop app search:

wme

The window opens. The blinking cursor is in a box at the top. Write away! Press Enter. The answer streams in below.

Want quick response?

wme -p "why is wryme's pet name wme?" --station "demo"

When you're done, press Ctrl-C to close it.

Keys you might want to know

Press	What happens
Enter	Send what you typed.
Esc	Stop a reply that's still coming in.
Ctrl-C	Close the program.
Ctrl-T	Switch between paged and scrolling view.
Ctrl-V	Read replies aloud on / off.
Ctrl-S	Open the station popup (a tabbed tuning menu).
F1	Open the popup directly on the Help tab.
Tab	In the popup: switch between Station / Help.
PgUp/PgDn	Scroll the popup body (also the mouse wheel).
arrow keys	Move the cursor inside your typing.
Backspace	Delete the character before the cursor.

Assistants talk to each other

Every open wryme shows its id in the status bar, like wryme#72e8. Ask one window to "ask wryme#72e8 what time it is" and its assistant sends a note and wryme#72e8 will reply back as guest@72e8. Web app assistants can participate if at least one native window is open.

What's cooking?

Wryme has both, web and desktop clients, besides TUI. All UX is consistent.

• For desktop, signed OSX releases are planned.
• For desktop, more multimedia support is planned.
• For web, memory and shell / machine use capability is planned. (besides through hive)
• hum hive support is set out to make the web version extremely powerful.

Trouble

• wme: command not found: your ~/.local/bin or ~/.cargo/bin isn't on your PATH. See the install section above.
• The status bar says station: demo: you haven't set up a shop yet. See "Set up your first shop" above.
• The status bar says unavailable: X: no shop lists station X's model, so wryme started on the first station that works instead. Add the model to a shop's models list or fix the spelling. Asking for X by name (--station X) fails with station 'X' wants model 'Y' but no shop advertises it.
• In the browser, a local shop's models are missing: a page can only read a local server that allows cross-origin requests. The server must send Access-Control-Allow-Origin and answer OPTIONS preflight requests. Ollama allows local pages by default.
• upstream 401: the key for that shop is wrong, missing, or expired.
• upstream 404: the model name is wrong, or the shop URL is wrong.

License

MIT. See LICENSE.