Using Myellm
What is it for?
Quick questions that come up while you’re doing something else: a vim motion, an ffmpeg command, a definition, or a choice between tools. Myellm combines a prompt for your chosen model with live search of previous answers, so you can ask, recall, and get back to work in the same place.
Is this a chat app?
The default workflow is built around short questions and reusable answers. When a follow-up needs a real conversation, Breakout Chat opens a separate multi-turn thread from an answer or from scratch. Breakout threads are durable, but promoting one into a titled reusable note is not built yet.
Is this a coding agent?
No. It answers questions and remembers the answers. It doesn’t edit files, inspect your repository, or run a tool loop over your codebase.
Do I have to remember exactly what I asked?
No. Type a few words to search your history. Keyword search is built in;
setting up a local embedding model with mm model setup adds semantic search
for related wording. Select a result to open the saved answer immediately.
Why is the second answer free?
Submitting an exact repeat returns the saved answer without sending a new request to your model provider. Opening a search result also reads the answer locally. Neither needs paid inference.
Automatically reusing an answer for a similar question is opt-in. That feature uses a conservative matching threshold and shows a similarity score. You can request a fresh answer if the match doesn’t help.
Do saved answers improve automatically?
Not yet. Today you can reuse saved answers and request fresh ones. Background distillation, which would improve a shared note using accumulated responses, is planned .
Models and cost
Can I choose the model?
Yes. mm --setup model saves the default model and reasoning effort, and each
profile can have its own. In the TUI, Alt-M and Alt-Shift-M cycle the model
for the next question; Alt-E and Alt-Shift-E cycle its effort. Those picks
last for the current TUI session without rewriting configuration. The refine
composer has separate model and effort controls. One model still answers at a
time.
Can I use a local model?
Yes. On Linux x86-64, mm --setup can explicitly download a pinned
Ministral 3 3B model and llama.cpp runtime—about 2.2 GB on disk and about
4 GB of RAM while loaded. Nothing downloads automatically, and installing it
does not replace a cloud default. myellmd runs it over a private Unix socket
and unloads it after idle time. You can instead point the managed runtime at
your own GGUF, or connect any OpenAI-compatible server you already run.
What does it cost?
Myellm itself currently charges nothing. New cloud answers use your own provider API keys and are billed by that provider; the opt-in managed model has no provider API charge. Saved answers cost nothing to retrieve. There is no built-in API spending limit yet.
Can I use a ChatGPT Plus or Claude Pro subscription?
Myellm connects using provider API keys. It doesn’t sign in with a ChatGPT or Claude consumer subscription. Set up API access and billing with the provider you want to use.
Can I ask several models at once?
Not yet. The planned workflow shows a fast answer while other models keep working, so a second opinion is already underway when you want it. Fresh questions currently dispatch to one selected model.
Your data
Where are my answers stored?
On your machine, in SQLite files under your XDG state directory. Those files use owner-only permissions. Your saved answers remain searchable and copyable offline.
Does it send telemetry?
Myellm records usage events locally for your own history and diagnostics. It doesn’t upload those logs. Cloud questions and configured context go to the selected provider; questions answered by the managed on-device model stay on your machine.
What do “private” and “local” mean?
Private limits a question to providers you’ve marked trusted. Ctrl+Enter uses this mode. By default, private providers must also run locally, though you can configure a trusted remote provider.
Local-only requires inference and embedding work to stay on this device. Private and local-only history lives in a separate store excluded from the planned sync path.
Can I undo a deletion or title change?
Yes. Deletion hides an answer without permanently erasing it, and manual or
generated title changes use the same durable action history. Press u outside
the prompt to undo the latest action and Ctrl-R to redo it.
mm --recover can also find and restore deleted answers after a restart.
Does my history follow me between machines?
Not yet. Each installation has its own local history. Multi-device sync and connecting to a daemon on another machine are on the roadmap.
Getting it
Can I try it now?
Not publicly yet. Myellm is in private daily use on its author’s machines. There is no public release date. The roadmap shows what’s next; the changelog records released builds.
Is it open source?
Not currently. The repository is private and unlicensed. Licensing and packaging for a public release are still undecided.
How is it installed today?
With source access, install.sh installs it for your user, or an Arch
PKGBUILD installs it system-wide. There’s no AUR package. mm --setup guides
cloud keys and endpoints, the default model and effort, profiles and notes,
hotkeys, the Omarchy bar icon, appearance, and advanced settings; mm --doctor checks provider credentials and config paths without changing them.
Can I test it without API keys?
With source access, a disposable Podman demo runs the real terminal UI against a fake provider without credentials or network access. It demonstrates the workflow; the answers are explicitly labeled fake.
Which platforms does it support?
The source builds for Linux and macOS. Private daily use and hardware verification currently center on Linux with Hyprland/Omarchy. The managed on-device model is Linux x86-64 only. The native Omarchy bar plugin requires Omarchy Quattro, and the standalone Wayland widget requires a wlr-layer-shell compositor. Sway, niri, and terminal-native macOS activation paths are implemented but not verified on their target hardware. Windows isn’t supported today.
Do I need Hyprland?
No. You can run mm in any supported terminal. Hyprland provides the
hardware-verified resident Super+M workflow; the standalone widget works on
supported layer-shell compositors, and the native bar panel is Omarchy-specific.
The project
Who built it?
Michael E. Gruen , for the questions he kept re-asking while learning new tools.
Why the name?
Myellm echoes “myelin,” the insulating sheath around nerve fibers that helps signals travel faster. The name nods to how repetition strengthens neural pathways: the more you return to something, the easier it becomes to recall. Say “my-el-el-em”: “my,” then the letters L, L, M.
Why a terminal app with a daemon?
The terminal keeps questions close to the work. A small Go client handles input and display while the daemon manages models and local memory. That split keeps typing responsive during inference and leaves room for other clients later.