Bring Your Own Agent
Connect any coding agent — Claude Code, Cursor, Codex — to drive Busabase. One prompt to onboard, one command to install a permanent skill.
Busabase has no built-in model. Instead it speaks plain HTTP, so you point your own agent at it — Claude Code, Cursor, Codex, Gemini CLI, or anything that can call a REST API. The agent writes through Change Requests, so every change carries a message, a diff, and an undo. With a proposal-only key, you stay the one who approves and merges.
You connect in two moves: paste one prompt to onboard, then install a permanent skill so the agent just knows Busabase every session.
Dashboard → Agent Skills: copy the prompt (it's localized to your UI language), or grab the MCP / OpenAPI endpoints.
Onboard with one prompt
Copy the prompt and paste it into your agent
Open Agent Skills in the dashboard sidebar and copy the short prompt. It points your agent at this workspace's live SETUP_SKILL.md — a self-contained onboarding script — and tells it to reply in your language.
The agent welcomes you and asks what to manage
Instead of dumping commands, it opens with a short welcome — what Busabase is, and how every change it makes is recorded and can be undone — and asks "what do you want to manage?" Answer in a sentence (e.g. "my sales leads / a CRM").
It connects and saves your config
The agent finds or creates an API key, verifies it can read your workspace, and writes the base URL + key to ~/.busabase/.env so every future session — and the installed skill — can read it. (See API keys to create a key yourself.)
It sketches the structure, you confirm, then it builds
Before creating anything, the agent shows you the planned shape — folders, bases, fields, relations — and waits for your go-ahead. On approval it creates the bases (always more than one node, so the workspace never opens empty) and you can see them in Graph View.
It seeds sample records and checks the write results
The agent writes a few starter records through Change Requests (never edited in place), so each one already carries its message and history. It checks the write responses to confirm the records landed, and reads a specific target only when required information is missing or the outcome is unclear. Want to see the review loop too? Ask it to send one change for review, then approve and merge it from the Inbox.
Install the skill for everyday use
Onboarding ends by installing a small, permanent busabase skill, so you never re-paste the prompt again:
npx skills add busabase/skillsThe generic skills CLI detects your agent (Claude Code, Cursor, …) and installs it globally. The skill is thin and evergreen — it reads ~/.busabase/.env and points at your live API for the full surface, so it never goes stale.
Setting up a brand-new workspace? The Quick Start walks you through the UI-first version (sign up → connect your agent → your first records) in ~5 minutes. This page is for handing the wheel to an agent.
Keep it up to date
busabase-cli, the Claude Code / Codex plugins, and the busabase/skills skill each update independently — there's no single "update everything" command. Instead of memorizing three different update commands, paste this into your agent and let it figure out what you have installed:
Check what Busabase tooling I have installed and update all of it to the latest version:
- If busabase-cli is installed globally (npm ls -g busabase-cli), reinstall it with
`npm i -g busabase-cli@latest`. If I only ever use `npx busabase-cli`, skip this —
npx already runs the latest release every time.
- If the Claude Code plugin is installed (`claude plugin list` shows busabase), run
`claude plugin marketplace update busabase` then `claude plugin update busabase@busabase`.
- If the Codex plugin is installed (`codex plugin list` shows busabase), run
`codex plugin marketplace upgrade busabase`.
- If the busabase / busabase-app-creator skill was installed with
`npx skills add busabase/skills` (check wherever this agent stores skills), re-run
that command to refresh it.
Tell me what you found and what you updated.Everyday: three ways to talk to it
Once connected, pick whichever fits the task — all three read the same ~/.busabase/.env:
| Method | Best for |
|---|---|
busabase-cli | The ergonomic everyday loop: npx busabase-cli bases list, change-requests review --change-request-id <id> --verdict approved, change-requests merge --change-request-id <id>. |
curl | Quick, zero-install calls against /api/v1. |
| MCP / OpenAPI | The complete, current surface — connect over MCP, or read /api/v1/openapi.json. |
The everyday loop is inspect the target → write a Change Request → check the write response. The response says whether the change landed or is waiting for review. Once it confirms the task succeeded, the agent can finish without another read, even when the response is a short acknowledgement or a merged ChangeRequest.
An extra read is useful when the response lacks information needed for the task, a timeout leaves the outcome uncertain, an asynchronous operation needs to finish, or you request additional verification. Appending attachments normally requires checking their returned references and metadata, not downloading the files.
See also: Core Concepts · Quick Start · AI Comments & @AI · REST API