Get started with the CLI
The terminal route — install the dibbla CLI, run dibbla init, and let your code agent drive it.
If you use Claude, Cowork, ChatGPT, Claude Code or Codex, three
steps connect it to Dibbla with nothing to install. This page is
the terminal route: the dibbla CLI, which ships code and manages
applications, databases, secrets and workflows from your own machine, and the
skill that teaches a code agent to drive it.
Pick the install method for your platform — all install the same binary.
Install
# macOS / Linux $curl https://install.dibbla.com -fsS | sh
# Windows $irm https://install.dibbla.com/install.ps1 | iex
# macOS / Linux $brew install dibbla-agents/tap/dibbla
All three methods land the same binary, and all three verify it before it reaches your disk.
You can also grab a binary directly from GitHub Releases,
which is where the .deb/.rpm packages and every previous version live.
Where the download comes from
The script installers talk to install.dibbla.com and nothing else. The version
list, the archive and the checksum file all come from that one origin, and the
archive’s SHA-256 is checked before anything is written to disk — if it doesn’t
match, the install stops and leaves nothing behind.
If you are asking an assistant — Claude Cowork, Claude Code, Cursor — to set Dibbla up for you, the command above is the one it should run, and it works from inside those sandboxes.
That is worth stating because it did not always. Agent sandboxes typically cannot
reach GitHub, so an installer that fetched its binary from GitHub Releases failed
there with Failed to fetch latest version. — nothing to do with your setup.
Dibbla now serves the binaries itself, so there is no GitHub call left in the
install path.
One asymmetry remains: dibbla update --version <tag>, which pins a specific
older release, still needs GitHub, because only the newest release is served from
install.dibbla.com. Plain dibbla update does not.
First-time setup
Run dibbla init inside any project directory. It’s a one-shot wrapper that gets you all the way to “shipping” in three guided steps — updating the binary, signing in, and installing the AI-agent skill so your editor can drive Dibbla for you.
$dibbla init ✦ Step 1/3: Updating dibbla to the latest version already on latest (1.2.33) ✦ Step 2/3: Logging in to Dibbla ? How would you like to log in? ▸ Log in with browser Paste an API token 🌐 Opening browser for login... ✅ Logged in to https://api.dibbla.com ✦ Step 3/3: Installing dibbla skill into current project writing .claude/skills/dibbla/SKILL.md writing .claude/skills/dibbla/guardrails.md writing .claude/skills/dibbla/user-docs.md … ten files in total updated AGENTS.md updated GEMINI.md ✓ Setup complete. Open your AI code agent in this folder and ask it to build and deploy your app.
Each step also runs as a standalone command — dibbla update, dibbla login, dibbla skills install dibbla. The login step is skipped if you already have
a token; pass --re-login to force it. The skill step is best-effort — if it
fails, the CLI continues. Use --user to install the skill into $HOME
instead of the current project (one install, every project on this machine).
Set up your code agent
dibbla init’s third step installs the Dibbla skill — a small bundle of Markdown that teaches your AI coding agent how to drive dibbla on your behalf. Pick your agent below to see what changed in your project and how to use it.
Claude Code reads skills from .claude/skills/. After dibbla init, your project has:
.claude/skills/dibbla/
├── SKILL.md ← entry point Claude Code reads first
├── guardrails.md ← the pre-deploy checklist, and when to ask first
├── user-docs.md ← how to write the handbook a deploy requires
├── examples.md ← copy-pasteable workflows
├── reference.md ← full command surface
├── manifest.md ← dibbla.yaml, for multi-service apps
├── platform.md ← what the platform gives a running app
├── workflows.md ← workflows and pipelines
├── sdk-go.md ← the Go SDK for workers
└── ai-gateway.md ← calling models without provider keysOpen the project in Claude Code and describe what you want built. The skill is what turns “deploy this” into the right sequence of commands:
Read the dibbla skill. Build me a small app that tracks which of our
internal tools are down, backed by a Dibbla Postgres database, then
prepare it for deploy and deploy it.Claude will pick up the skill on the next session — no extra wiring.
Cursor reads AGENTS.md at the project root (open standard). dibbla init either created that file or merged a marker-delimited block into your existing one — your own content is preserved.
Restart your Cursor session (or open a new chat) and try:
Use the dibbla skill to deploy this directory as a new app.Cursor’s agent will load the AGENTS.md context and use the skill’s reference + examples.
Gemini CLI reads GEMINI.md at the project root. dibbla init wrote (or merged into) that file with the skill’s full context.
Run gemini inside your project and try:
Use the dibbla skill to deploy this directory as a new app.Gemini will index GEMINI.md on startup and follow the skill’s instructions.
Opencode reads AGENTS.md at the project root (open standard). dibbla init either created that file or merged a marker-delimited block into your existing one — your own content is preserved.
Restart your Opencode session and try:
Use the dibbla skill to deploy this directory as a new app.Most modern coding agents — Codex, Copilot, Windsurf, Aider, Zed, Warp, RooCode — read AGENTS.md at the project root (open standard). dibbla init wrote (or merged into) that file with the skill’s full context, so any AGENTS.md-aware tool picks it up automatically.
If your agent doesn’t read AGENTS.md yet, paste the output of dibbla --skill-prompt into its system prompt or context window — it contains the same instructions written for direct LLM consumption.
See the skills guide for per-project vs $HOME install, the bundled skill list, and troubleshooting.
No terminal? Connect an assistant instead
The CLI is one way in. The other is the Dibbla connector: Claude, Cowork, ChatGPT, Claude Code or Codex connects to Dibbla over one OAuth-protected URL and reads, changes and deploys your apps for you. Nothing is installed on your machine and no token is pasted anywhere.
- Connect your assistant — three steps, per assistant.
- Disconnect and revoke access — cut a connected assistant off yourself, at any time, without contacting support.
Next
Three steps from Claude, Cowork, ChatGPT, Claude Code or Codex to a connection that can deploy — and how to revoke it again.
Build one real app from nothing to a nightly pipeline, step by step.
Make Claude Code, Cursor, Gemini CLI, and other agents fluent in Dibbla.
Every command, flag, environment variable, and shell-completion script.