Koder · desktop 0.4.53
Koder is a tool for coding and office work. It runs on your computer and does not include a model: you set up the model service and the key. It reads and writes files, and runs commands, only in the folder you pick. It asks before changing a file, and by default it asks before running a command.
IDECLITUIThe windowTutorialSettingsApprovalsWorkbenchComputer useModelsDetailsData
The IDE is in the desktop installer, and its window has its own terminal at the bottom. The CLI and the TUI are each one command, on the Koder CLI page. The install steps are in the project repository. The TUI is not in this installer.
The dev view is the editor: the file tree on the left, the open file in the middle, and this turn’s chat on the right. The work view puts the chat in the middle. A file change shows the before and after, and it is written only after you accept it. The terminal is at the bottom, opened with ⌘J or Ctrl+J. See the IDE
The command is koder. It is not in the desktop installer and needs no invite code. The install steps are in the project repository. Start a session or give a task in one line, and use it in scripts and CI too. A file change is listed first and written only when you type y. Commands ask first too. See the CLI
A full-screen terminal interface: the chat, a side column, and a prompt. A file change opens a diff first. Press y to accept, or n or Esc to turn it down. Commands ask first. It does not collect passwords. It is not in this installer. The install steps are in the project repository. See the TUI
The window
Tasks, the workbench, and Settings share that window. Below is how the current app is laid out: an activity bar on the left, a list beside it, and the work in the middle.
The top of the task window switches between a work view and a dev view. In the dev view, tasks and files are on the left, the editor is in the middle, and this turn’s chat is on the right. The work view puts the chat in the middle. When a file needs to change, the chat shows the before and after first. It is written only when you click Accept.
It sits next to Koder tasks and KARMEN bot. The home screen is a row of shortcuts, three first-run steps, counts of projects and work, and today’s pulse. All of that stays on this computer.
Four groups on the left, and you can search. Approval modes are Careful, Balanced, and Automatic. What each page does is on Settings.
The workbench sits next to tasks and KARMEN bot. The home screen shows goal progress, tasks in progress, the board, recent work, and the next automation. Workbench data lives in the Koder data folder on your computer.
Each project is a folder, with a board of To do, In progress, and Done. Click Hand to AI on a card and a task opens in that project’s folder.
Images, web pages, documents, and audio or video in the project are listed for you. Files the AI wrote are marked. Click one to preview it.
Add text, Markdown, web pages, Word, and PDF. Text inside images can be read too. The index is built on this computer. AI searches only the libraries you turned on, by keyword.
Collect files and links, with notes and tags. Send one into the knowledge base in one step, or copy it into a project for a task.
Draw a few prompts as a flow. Drag boxes, connect them, and add an If branch so a yes and a no take different paths. When it runs, one task follows the arrows to the end.
Run a prompt or a workflow every day, every week, every few hours, or once. It runs only while Koder is open. A missed run is skipped and noted by default. You can also check Catch up on missed runs, and it runs the next time you open Koder.
Write a goal and the key results, check them off, and keep the tasks you opened for it together. A single task can have a goal too. AI pushes it round by round, and it pauses, with a reason, if something fails or it hits the round limit.
Turn on a lead for a project. It can see the other tasks in that project, leave them notes, and stop a task that has drifted. It also keeps the project rules and the log. It does not run on its own. It works only when you open it.
Ask up to three models the same question and compare them side by side. You can also try a knowledge-base search, or check whether a command would run without asking. Comparisons call the providers, and those calls cost money under their rules.
Preview
This version does not include a model, and it does not bill you. You connect your own model service in Settings. ESVORIQ 1.0 is in training. Public weights are not available yet. Until then, the tools below help you use the service you already have.
Nearly thirty model services are preset, in China and elsewhere, and you can connect a local model or another compatible endpoint. Each one says where to get a key. The first time you open Koder, it walks you through connecting the first model.
This version has no Lite, Pro, Max, or Ultra strength buttons next to the input. The hourglass picks a model for simple, medium, and complex work, and each level’s thinking can be Default, Low, Medium, or High.
When it is on, the first line of each new task goes to a fast, cheaper probe model that judges the difficulty, then to the model you set for small, medium, or large. The top of the task says which level it chose, and why. It is off by default.
After you add a key, you can read the provider’s live model list. You can also point one provider at a different endpoint, or add a model that is not on the list.
1 / 2
While AI is still working, a new message can wait in line and go out when this round ends. Or you can send it now, and AI sees it on the next step.
When AI needs you to decide, it shows a card of choices. If you do not answer in 60 seconds, it continues on its own judgment, and you can still add an answer later.
Pin a task, a chat, or a folder. Koder can remind you when a task finishes or needs you, send a system notification, or show an icon in the menu bar.
If a task is still running when you quit, Koder warns you first. The next time you open it, an interrupted task can continue in one click.
The app has more than forty illustrated tutorials. A few of them move the pointer and walk you through the real interface. The steps on this site are in the tutorial. What each settings page does is in Settings.
Light, dark, tint, font and size, high contrast, and reduced transparency. Put the sidebar on the left or the right, and hide the activity bar. The dev view has an editor and a terminal, and the chat can sit on either side of the editor.
Koder runs on your computer. We do not collect what you write. What you send to a model service is handled by the provider you chose.
On our side
The model service you chose
That is how 0.4.53 works today. If it changes, the privacy notice is updated first, and the app tells you.
The interface is served by a local service that listens only on 127.0.0.1, and it refuses requests that come from other websites.
You open the model service, get the key, and pay that provider. What you can connect is the list inside the app, and you can also connect a local or other compatible endpoint.
API keys are stored with AES-256-GCM. The key file used for that encryption can be read only by the current system user. Koder does not read model keys from environment variables or from a .env file, and the interface does not show any part of a key. It is not in the macOS Keychain yet.
The System page shows token use and an estimated cost, plus the version and the security patch level, and you can run a check in one click. A usage record is the time of the call, the task and folder, the service and model, and the input and output token counts. It does not include the message or the file, and it is not uploaded.
KARMEN LOOP is a kernel we wrote ourselves. It finishes a task in a loop of perceive, decide, and act. It owns the session, the main loop, tool calls, and approvals. When a file needs to change, it stops and waits for you. Commands follow the rule you set, and by default they ask too. If Koder quits unexpectedly, an interrupted task can continue from where it stopped.
Model calls, files and commands, web pages, and computer use each run in their own subprocess. If one of those exits with an error, the kernel keeps running and starts a new one the next time it is needed.
Koder for phone is in development. There is no installer, and the download page has no phone install button.