Early access

One agent helps. A crew ships.

One app for all your AI coding agents. Bring Claude Code, Codex or the built-in engine with your own plans and keys, give a leader the goal, and a crew builds it side by side while you watch every step. Free for macOS and Windows, in private beta, and it runs on your computer.

Get an invite Sign in

Local is local. Your chats, files, keys and memory stay on your computer. No tracking.

Private by design

Local is local.

Yami is an app on your computer, not a website and not a cloud. Your work lives where you do. Here is exactly what stays there, and what leaves only when you use it.

Stays on your computer

  • Your chatsEvery message and its history, in the app's own folder. Put a password on a chat and Yami's copy of it is encrypted.
  • Your filesAgents work in your folders on this computer, each member of a crew in its own git worktree.
  • Your keys and sign-insThe keys you give Yami are sealed with a key only your user account can read, and never sent to Yami. Claude Code, Codex and the built-in engine keep their own sign-ins.
  • What Yami remembersIts memory of you and of each project, as notes you can read and change.
  • Everything elseNotes, drafts, routines, settings, token counts and the code map.

Leaves only when you use it

  • The engine you pickWhat the agent needs for an answer goes to that engine's model, as it would without Yami: DeepSeek for the built-in engine (its servers are in China), Anthropic for Claude Code, OpenAI for Codex, or the service you add a key for. A model running on your own computer sends nothing.
  • The web, when askedPages an agent reads, sites you open in the built-in browser, pictures in a reply.
  • Add-ons you connectEach talks to its own service. Browsing the store searches the public MCP Registry and GitHub.
  • RemoteYour phone and your computer meet on your own Tailscale network, never at a public address. Notifications reach your phone through its own push service, encrypted.

And Yami itself? Almost nothing: your beta sign-in and a check that it is still good, a look for a new version, and feedback only when you send it. No analytics, no tracking, no ads. Every detail.

The reel

One ask. Merged work.

  1. You askIn plain words, the whole thing.
  2. The leader plansParts you can read, and a question before it spends.
  3. A crew startsThree agents, each in its own worktree.
  4. You see whereThe map lights the files they are in.
  5. Yami mergesEach branch into the one you are on, on your OK unless you let the crew run.
  6. DoneThe leader wakes and tells you. The cat is pleased.
LeaderNight Paws
  • Save slots in save/store.js
  • Level 3 in levels.js, with tests
  • Menus for the new saves
Start Sleepy Miso? StartDon't start
Sleepy MisoWriting the save slots
Velvet PepperDrawing level 3
Tiny MochiLaying out the menus
Merged Sleepy MisoMerged Velvet PepperMerged Tiny Mochi
Done · Ship saves and level 3

Not a chatbot. A crew.

01 · Ventures

Hand it the whole thing.

A launch, a game, a company site with its brand. A leader agent writes the plan and puts a crew to work on the parts at the same time. You choose: ask before each agent, or let the crew run.

  1. PlanThe leader breaks the goal into parts you can read.
  2. AskYour approval mode decides: a card before each agent starts, or the crew starts on its own.
  3. CrewUp to eight at once (four unless you change it), each in its own git worktree, never in each other's files.
  4. MergeWhen the leader says a branch is good, Yami merges it into the branch you are on. It asks you first unless you chose Bypass, which merges, logs it and offers an Undo.
  5. WakeThe leader wakes by itself when the crew is done, and tells you.
8agents side by side, at most, by your setting
1worktree each, so no two agents touch the same checkout
0agents start without your OK in Manual. Auto and Bypass let the crew run.
Youset the token budget. At the limit, no new agent starts until you raise it.

02 · Tokens

Every token, in plain sight.

Every chat has a card with its bill: how much was new, how much came from the cache and, for Claude Code, Codex and the built-in engine, about how much Yami itself added.

Chat stats
67.9Ktokens in this chat
Built-in engine67.9K80% cached
14.1K new tokens (the rest came from the cache)~0.9K added by Yami (its notes, memory, venture)

Every chat shows its bill.

Yami adds its notes to the agent when a chat starts, and again only when they change. A card in every chat says how much was new, how much came from the cache, and about how much Yami added. A venture can have a budget: at the limit, no new agent starts.

03 · See it work

Nothing happens in the dark.

A line that says what it is doing right now. A map of your code where the files an agent is in light up. And a cat on your desktop that tells you, frowns while it works, and gets heart eyes when it is done.

JavaScriptHTMLCSSan agent is here
Thinking“Checking how the save slots load”
Working“Running npm test · 3 steps”
Done“Done · Ship saves and level 3”

04 · Looks

Night or day. Loud or quiet.

Yami and Yami White carry the aurora and the glow. Clean and Clean White keep to greys, with no aurora and no intro, for days you want it plain. Pick one, and it spreads across the window.

Yami's start screen in the Yami look: the violet cat on a night sky, the message field under it. The same screen in Yami White. The same screen in Clean: greys only, a white cat. The same screen in Clean White: white and greys, a dark cat.

05 · Every agent

One window, more than one engine.

Bring your own plans and keys. Under the hood Yami runs the first-party Claude Code and Codex, its built-in engine, or any key you add: one window, one history, the same app tools for each.

If installed

Claude Code

Have Claude Code on your computer? Yami can run it in a chat, with your linked folders and the app's own tools.

If installed

Codex

Have the Codex CLI? Yami can run it too, in the same calm window.

Comes with Yami

Built in

Nothing to install: sign in with a DeepSeek account or add a key, and it works. You choose how hard it thinks.

Why Yami

Agents are smart. Yami makes them a team.

Yami is not another model. It is where your agents work together: it plans, splits the work, shows every step and counts every token, whichever engine you use.

Big jobsOn its ownYou split the work and paste each part inWith YamiA leader plans it, and you choose if each agent waits for your OK
Many handsOn its ownOne agent in one checkoutWith YamiUp to eight at once, each in its own git worktree, merged on your OK or on their own in Bypass
Seeing it workOn its ownText scrolling byWith YamiA live line, a code map that lights up, a cat on your desktop
The billOn its ownEach tool its own command, in its own formWith YamiOne card per chat, any engine: new and cached tokens; ventures can have budgets
EnginesOn its ownThe one it was built forWith YamiThe built-in engine, and Claude Code or Codex if you have them, per chat
Away from the deskOn its ownCome back and read the scrollWith YamiApprove from your phone over your own Tailscale
Keeps goingOn its ownStops when the answer endsWith Yami/goal works turn after turn; the leader wakes when its crew is done

06 · Everything else

The rest of the toolbox.

Each one is in the app today. Scroll, and each plays as you reach it.

  1. /goal

    Give a chat a goal. It works turn after turn until the goal is met, or until it needs you.

  2. Routines

    Every morning, every hour: work that runs on its own while Yami is open, even in the tray.

  3. Diagrams

    43 kinds, drawn live in the chat and editable.

  4. Images

    Ask for one and your ChatGPT plan draws it, with the brief kept beside it.

  5. Remote

    Your phone steers your computer over your own Tailscale. Swipe to approve. Face ID to open. Never a public address.

  6. Browser and terminal

    Built in, with your sign-ins. The agent uses them too.

  7. Approvals

    Auto, Manual, Accept edits, Plan or Bypass permissions. The built-in engine and Claude Code stop at the same approval card; Codex keeps its own sandbox. The same mode decides if a venture's agents start and merge on your OK: Auto starts them, Bypass also merges.

  8. Thousands of add-ons

    From the official MCP Registry, searchable on your computer. GitHub with one sign-in.

  9. Import chats

    Bring your chats in from Claude Code, Codex, OpenCode, Gemini CLI, Cline, Roo Code and Aider, or from your ChatGPT and Claude.ai export. Claude Code and Codex chats go on where they left off. Read only, and nothing leaves your computer.

  10. Memory and skills

    It remembers what you tell it, and learns what you teach it.

  11. Steer or queue

    Write while it works: change its course now, or line the next task up.

  12. Quick chat

    A key from anywhere on your computer, and Yami is there.

  13. Fork a reply

    Try another way from any point, and keep both.

  14. Four looks

    Yami, Yami White, Clean and Clean White. See them change.

Yami costs$0No tiers. No ads.
  • Every feature on this page, with no tier above it
  • A built-in engine, and Claude Code and Codex if you have them
  • Runs on your computer. Your files stay there.
  • You pay only for the model you use, with your own plan or key

07 · Questions

Asked often.

Is it really free?

Yes. Yami itself costs nothing. The models cost what their provider charges, as they would without Yami.

Does my code leave my computer?

Yami runs on your machine and keeps your chats there. The engine you pick sends what it needs to its model, as it would without Yami.

How is this different from using Claude Code or Codex alone?

Yami is an app around agents: ventures with a crew working in their own worktrees, a code map, a cat that tells you what is going on, and a bill you can read. It can run those tools for you if you have them installed.

Can agents break each other's work?

No two agents work in the same checkout. Each has its own git worktree and branch, and Yami merges a branch into yours on your OK, or by itself in Bypass, logged and with an Undo. A folder without git gets one agent at a time.

Which systems?

macOS on Apple silicon and Windows 11 (x64). Early access opens in waves.

How does it work, under the hood?
  • A desktop app. Yami runs on your computer. Your chats, files, memory and keys stay there; keys are sealed with a key only your user account can read.
  • Engines. The built-in engine (DeepSeek Harness), your own Claude Code or Codex on your own plan, an API key for Anthropic or an OpenAI-compatible service, or a model running on your own computer.
  • Ventures. A leader agent writes the plan, up to eight agents work on the parts at once, each in its own git worktree and branch, and Yami merges their work into yours.
  • Approvals. You pick how much it may do alone: Auto, Manual, Accept edits, Plan or Bypass. Typing, clicking or opening apps on your computer always asks first.
  • Your phone. Yami Remote runs over your own Tailscale network, with Face ID to open. It never has a public address.
  • Updates. Every update is signed, and Yami installs only one whose signature checks out.
What does Yami itself send anywhere?

Very little: the beta sign-in and a check that it is still good, a look for a new version, and feedback only when you send it. No analytics, no tracking, no ads. What your engine sends to its model is between you and that provider. Every connection is listed on the privacy page.

Why does macOS or Windows warn me when I open it?

Yami is a free project made by one person, and Apple and Microsoft charge every year for the certificates that silence that warning, so the installers are not notarized by Apple or signed for Windows. The first-open steps in the download window take you past it once. Every update after that is signed with Yami's own key, and Yami installs only an update whose signature checks out. The download window also shows the installer's SHA-256, so you can check the file you got is the one published.

Why isn't Yami open source?

Yami is free to use, but its code is mine to keep. It is a one-person project, and I would rather spend my time making it better than running an open-source project around it. It stands on open source, OpenGhost and DSH (both MIT), and credits them.

Who made it?

Tupe, one person. Yami started as a hobby. I was tired of jumping between Claude Code, Codex, chat apps and my own projects, so I built one place for all of them, with a crew of agents on top. It turned out good enough that I wanted to share it. It is free, with no ads and nothing sold, and I pay for the servers myself.

Yami is built on OpenGhost, the MIT-licensed desktop agent by Andre (@ANDRETRIPOL on GitHub).

Your crew is waiting. Yami is in private beta, and invites go out in waves.

Get an invite Sign in