:::::::..    .,::::::    .        :     .-:.     ::-.;;;;``;;;;   ;;;;''''    ;;,.    ;;;     ';;.   ;;;;' [[[,/[[['    [[cccc     [[[[, ,[[[[,      '[[,[[[' $$$$$$c      $$""""     $$$$$$$$"$$$        c$$" 888b "88bo,  888oo,__   888 Y88" 888o     ,8P"` MMMM   "W"   """"YUMMM  MMM  M'  "MMM    mM"

Real work, done wherever you are.

Remy is one calm assistant for real digital work. Describe a task in plain words; it reads, searches and acts in small steps, and asks before anything that matters.

It lives in your terminal today. Web, desktop, mobile and voice are on the way.

Get Remy Follow along on GitHub

Terminal · available now

$ npm install -g @koryei/remy

Needs Node 22.18 or newer. Then run remy to open the chat. Model setup is below.

Reach Remy anywhere.

The same assistant, wherever you already are. One surface exists today. The rest are planned, in this order, and have no dates yet.

Terminal
Available now
Open a chat, describe the task, approve what matters. Works with the model you choose.
Web
Next
The same chat in a browser, with a live timeline and approval cards. Nothing to install.
Desktop
Later
Remy on your own machine, working with your apps only where you allow it.
Mobile
Later
Start a task, get a note when it needs a yes, and approve from your pocket.
Voice
Later
Say what you need. Remy answers briefly and still asks before anything that matters.

One task, continued across surfaces. A preview of the intent. Only the terminal step exists today, and it stops at the saved draft. Sending needs an email integration that is not built yet, so the phone and voice steps are mockups.

Terminal · works today

you Compare the quotes in quotes/ and draft a reply to Dana ✓ read_file quotes/*.md ✓ write_file reply.md Draft saved to reply.md. Nothing is sent.
Mobile · preview

Remy needs a yes Send reply.md to Dana. Approval covers this exact message.
Voice · preview

you Send it. ✓ Sent to Dana. 4 steps, $0.0009.

What it does today.

In the terminal, Remy works in small steps with these tools until the task is done, it needs you, or it hits a limit.

FilesRead, list, write, edit, move and delete inside one workspace folder.
WebFetch pages. Search with a Brave, Tavily or SearXNG key. Limit browsing to domains you name.
CommandsRun one allowlisted program at a time, such as git, grep, node or pnpm. Every command asks first.
MathArithmetic and date math, such as the days between two dates.
QuestionsWhen it needs something only you know, it asks. At most three times per task.
ModelsAny OpenAI-compatible endpoint: OpenAI, OpenRouter, Ollama, vLLM or Groq.

you  Summarize https://example.com/pricing into notes.md
Plan: fetch the page, write the summary.
✓ fetch_url  example.com/pricing  1.2s
  Page content is untrusted. Writes now need approval.
Approve? write_file notes.md  (18 lines)  [y/n]
y
✓ write_file  notes.md
Saved notes.md. Three plans, priced per seat, with the limits of each. The page does not say whether annual billing is cheaper.
3 steps  ·  $0.0004
      

An example session in the terminal today. Remy fetches a page, then stops to ask before it writes a file.

Nothing happens that you did not approve.

Approvals bind to the exact action

Each yes covers one specific call and is used once. Deleting files, running commands and spending money always ask, whatever you allow.

Untrusted content is marked

Text from web pages and command output is untrusted. Once a task has seen it, every action that is not a read needs your approval, so a hidden instruction in a page cannot act for you.

Crash-safe resume

Every task is an append-only log. After a crash or Ctrl+C, remy resume continues from where it stopped.

No double actions

An action that may already have happened is checked or reported, never blindly repeated. Tasks stop at 30 steps, $1 or 30 minutes.

Measured, not claimed.

99%

169 of 170 runspassed: the 34-case suite, five runs each, on GPT-5 nano

0safety violations and duplicate actions

about $0.0008per task

Run 9, 2026-10-26: 99% success, 97% of cases passing all five runs. The one miss asked too many questions and never sent. The earlier 30-case suite scored 100%. Weaker or free models are less reliable than this; open-weight baselines are still to be measured. Full history is in evals/results.

Pick a model and talk.

Pick a model

export REMY_PROFILE=openrouter
export OPENROUTER_API_KEY=sk-or-...
export REMY_MODEL=google/gemma-4-26b-a4b-it:free

# or local
# REMY_PROFILE=ollama REMY_MODEL=qwen2.5:7b

Open the chat

remy

# one task, for scripts
remy run "summarize README.md"

Not installed yet? The commands are under Get Remy, at the top of the page.

Not yet.

Only the terminal exists. Web, desktop, mobile and voice are planned and not built.
No browser automation: it reads pages but cannot click or fill forms.
No real email, calendar or account integrations.
Ctrl+C stops the task, but a tool already running finishes its current call.
run_command is not a full sandbox. Read each command before you approve it.
No memory across chats beyond the current conversation, and no scheduled tasks.
Free and weaker models are less reliable than the numbers above.