You explain
In plain language, in the project conversation: what is wrong, or what you would like. No commands, no syntax - ordinary sentences.
Vibe coding, on your own server
Your vibe coding tool,
working while you sleep.
HaikoDev is an open-source vibe coding tool, sitting on your own server. You describe what you want, card by card: every card is born with its own agent, which discusses the need and then hands the work to a full agent, online. Switch off your computer: the work carries on without you, and you read what got done when you come back.
Loading the board…
What you see above is not a picture: it is the real HaikoDev interface, with made-up projects and made-up cards. Switch projects, move a card, or click “New task”: an agent is born with the card, discusses your need, writes the card — and you can start it.
What HaikoDev does
Vibe coding, taken out of the terminal and put on a board
You know the idea: you say what you want, the machine writes the code. HaikoDev adds what is missing the moment you actually want to make progress - a board of cards, several agents started together, and a server that works while you do something else.
An agent writes the card
A lightweight agent is born with the card, discusses your need, then writes the card in a fixed format: short title, goal, thinking level. Not a wall of free-form text to wade through.
The agents work
Several at once, each on its own branch and its own copy of the project, on your server. Close the tab: it carries on without you.
You decide
You read it over, drag the card into "To deploy", and publish whenever suits you. The batch goes out in one piece.
The five columns
Every column shows the agent’s role
Each column is one stage in an agent’s work: one glance tells you where each card stands, without reading a single label. And above all, who is allowed to move it.
Planned
Every card is born here. It carries a short request, a thinking level and, if you want, a start date. Nothing has gone to the engine yet: while the card sleeps, it costs nothing.
Running
An agent has taken the card. It has its own branch and its own copy of the project; it reads, edits, tests and shows you its task list ticking off live. A question from the agent stops everything until you answer.
Done
The agent has handed in its report: what it understood, what it did, the consequences, and the outcome of the checks - failures included. The card lands here on its own; nobody pushes it.
To deploy
You have read it over and you approve. Cards pile up here as one batch, ready to go out together. No agent walks through this door: publishing is a human gesture, always.
Archived
What is settled leaves the board without disappearing. The history, the branches, the reports and the estimates stay readable - and only reopen on your gesture.
Open sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens countedOpen sourceOn your own serverClaudeCodexCursorA memory that files itselfTokens counted
The features
What an AI assistant is missing when you hand it a real project
A model that writes code, you already have one. What changes everything is the frame around it: tasks carried through without you, answers always in the same shape, and tokens you do not burn for nothing.
A board, not a chat window
Your projects, your cards and your agents fit on a single screen. You drag a card from one column to the next; the work follows. Nothing to dig back out of an endless conversation.
One agent per card, writing the card rather than a wall of text
Every card is born with its own agent, on a low-cost model. You describe the need in plain language, it talks it over with you, then writes the card in a fixed format - title, goal, thinking level - and nothing starts until you press “Start the task”.
A plan mode before a single line is written
On delicate topics, the agent proposes a plan first, which you refine back and forth. Nothing starts until the plan suits you.
As many cards running at once as you like
Every agent works in its own copy of the project, on its own branch. Five cards started together move together, and never tread on each other.
The AI tools already on your server
HaikoDev detects the AI tools installed on your VPS. From the settings, you can add more, then use them all through the same interface.
A memory that files itself
Every agent that finishes drops off what it learned; the memory sorts itself by topic and grows with the project. When a card starts, only the useful passages go to the engine - not the whole folder.
You describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machineYou describe it, it gets builtSeveral tasks at the same timeYour computer can be switched offAnswers in a fixed formatYou publish, not the machine
The benefits
What HaikoDev does that a raw Claude or GPT does not
The model is the same. What changes is the frame: fixed formats, a memory that structures itself, tokens counted - and safeguards agents cannot remove.
You keep control of publishing
No agent ever publishes on its own initiative. Code gets written, committed and pushed by itself, but it waits in "To deploy" until you click. It is the hardest rule in the tool, and there is no way around it.
Answers in a fixed format, not free-form text
A general-purpose assistant answers whatever it likes, however it likes. Here every step returns a shape decided in advance: a card has a title, a goal and a level; a report has its sections, its consequences and its estimate. You always read the same thing, and you see straight away what is missing.
You follow the work without reading code
Every agent announces its task list and ticks it off in front of you, then hands back a report in plain language: what it understood, what it did, what it changes. Failed tests are written there too - a failed check is never hushed up.
You do not burn your tokens
A sleeping card consumes nothing. The project memory structures itself with every finished card: the next agent gets the three useful passages instead of the whole folder. And what goes to the engine is measured, card by card, not assumed.
Nothing gets stuck
A run that ends closes its card, a passing outage is retried, a run cut short by an account limit starts again, and a sweep catches forgotten cards. A card in "Running" always says what is still going on.
You stay at home, with the code
HaikoDev is open source and runs on your own server, with your accounts and your repositories. The whole project is handed to you: you read it, you change it, you keep it. Nothing depends on a service that could shut down tomorrow.
0
computers to leave switched on: it all runs on your server
3
engines drivable behind one interface: Claude, Codex, Cursor
7
board columns, five of which you handle every day
5
interface languages, set per project
The changelog
What has been shipped, in order
HaikoDev moves fast: every published card is added to a public timeline, day by day. Nothing is hidden, quiet days included.
- Mode conversation vocale : écoute au clic (bleue), réponse vocale, couper la parole
- Raccourci clavier réglable pour activer l'écoute vocale
- Conversation vocale : écoute au clic sur le module, réponse voix + texte, fil qui défile
- Bouton de déploiement : libellé de l'étape en cours + dropdown du process avec descriptions
- « Compte Max 20x en principal, Pro en secondaire » livrée.
Put it on your own server
Pull the project, install it on your server, plug in your accounts. The whole code is yours, and the guide skips no step.
Support the project
A donation to keep HaikoDev going
HaikoDev is an independent project, built in Switzerland. A donation, in Swiss francs or euros, helps fund the time spent on it.
Secure payment via Stripe, in CHF or EUR.