Agent Skills

An AI assistant that has read Rebase’s documentation writes better Rebase code than one guessing from the shape of the API. rebase skills install copies 21 Markdown skill files into your repository, in whatever layout your assistant expects:

rebase skills install

The skills are reference material, not tools. They tell an assistant how collections are defined, why migrations are two steps, and which mistakes the framework will not catch for it. For tools that act on your data, see the MCP server.

Set up by rebase init

A new project does not need the command. rebase init asks whether to set up your AI coding agents, then lists the ones it finds on the machine:

? Set up Rebase skills and the Rebase MCP server for your AI coding agent(s)? Yes

Detecting installed AI coding agents...
  ✓ Claude Code — ~/.claude
  ✓ Gemini CLI / Antigravity — ~/.gemini
  ✗ Cursor — ~/.cursor (not found)
  ✗ Windsurf — ~/.codeium/windsurf (not found)
  ✗ Codex CLI — ~/.codex (not found)
  ✗ Kiro — ~/.kiro (not found)
  · GitHub Copilot — can't be detected; tick it below if you use it

The agents it found are pre-ticked. For each one you keep, it writes the skills and registers the MCP server in that agent’s project config, all before the project’s first commit. In CI, or with --yes, name them:

rebase init my-app --yes --agent claude,cursor

An existing project

rebase init refuses a directory that already holds a project, so a project made before agent setup existed — or one whose author declined the prompt — gets the same setup from rebase skills install --mcp: the skills, and the MCP server registered in each agent’s project config. Servers already in the file are kept, and a rerun leaves the Rebase entry as it is.

rebase skills install --agent cursor --mcp

Without --mcp the command writes only the skills, and names the agents whose config does not have the server yet.

Which assistant

The command takes --agent (or -a), repeatable and comma-separated:

rebase skills install --agent claude
rebase skills install --agent claude,cursor
rebase skills install --agent all

Seven targets are supported — one for every pointer file rebase init writes:

--agent	Assistant	Written to
cursor	Cursor	.cursor/rules/rebase.mdc + .cursor/rules/<skill>/SKILL.md
claude	Claude Code	.claude/skills/<skill>/SKILL.md
windsurf	Windsurf	.windsurf/rules/rebase.md + .windsurf/rules/<skill>/SKILL.md
gemini	Gemini CLI / Antigravity	.agents/skills/<skill>/SKILL.md
codex	Codex CLI	.agents/skills/<skill>/SKILL.md
kiro	Kiro	.kiro/steering/rebase.md + .kiro/steering/<skill>/SKILL.md
copilot	GitHub Copilot	.github/instructions/rebase.instructions.md + <skill>/SKILL.md

Cursor, Windsurf, Kiro and Copilot get one always-on file

Those four load their whole rules directory into every request. A rule file per skill meant about 84,000 characters of Rebase reference in front of every question a person asked, whether or not it was about Rebase — and an instruction an assistant skims is one it does not follow.

They get rebase.mdc (or rebase.md) instead: a ~3 KB index with alwaysApply: true, listing what each skill covers and the file to read. The bodies sit in per-skill subdirectories and are opened on demand.

gemini covers both Gemini CLI and Antigravity — they read the same .agents/ directory, so there is no separate antigravity value. Codex reads it too; naming both gemini and codex writes the directory once.

With no --agent, the command detects which assistants a project already uses by looking for .cursor/, .claude/, .windsurf/, .agents/, .codex/ and .kiro/. If it finds none it prompts you to choose, with the assistants installed on the machine already ticked.

GitHub Copilot is never detected. Its directory would be .github/, and .github/ is not evidence anyone uses Copilot: rebase init writes .github/copilot-instructions.md into every scaffold, and most repositories have a .github/ for workflows. Install it with --agent copilot.

A freshly scaffolded project always prompts

rebase init writes CLAUDE.md, .cursorrules and friends, but none of the directories detection looks for. So the first run in a new project falls through to the prompt — and in CI, where there is no TTY, it exits with an error instead. Pass --agent explicitly in any non-interactive context.

Project-local, and meant to be committed

Skills are written relative to your project root — the nearest ancestor containing rebase.json — not to your home directory and not to the current working directory. Nothing is installed globally.

Commit them. They are part of the repo the same way a lint config is: every contributor’s assistant then works from the same understanding of the codebase, including contributors who never ran the command.

Re-run the command to update after a Rebase upgrade:

rebase skills install --agent all

Each target directory holds a .rebase-skills.json manifest of what the command wrote, so a re-run can tell an edit from an older release’s copy. Commit it with the skills.

• A file nobody edited is replaced with the new release’s copy.
• A file you edited is kept, and the command lists it. --force replaces it.
• A skill a newer release no longer ships is removed, unless you edited it. Then it is kept and listed.
• Files the command never wrote are not touched.

Project-specific guidance still belongs in ai-instructions.md, which is yours. An edited skill stops receiving updates until you --force it.

The command also works outside a Rebase project, falling back to the working directory — useful for a separate frontend repo that talks to a Rebase backend.

The 21 skills

Skill	Covers
rebase-basics	Core principles, workflow and maintenance — the entry point the others assume
rebase-collections	Defining collections, property types, validation, searchability
rebase-backend-postgres	The Postgres backend: setup, schema generation, migrations, pooling, read replicas
rebase-api	The generated REST API — endpoints, filtering, sorting, pagination
rebase-sdk	The generated TypeScript SDK: CRUD, filtering, search, auth, realtime, offline, storage
rebase-auth	Authentication, roles, RLS policies, MFA, API keys, OAuth, custom adapters
rebase-security	Access control, interception, fail-closed design, PII masking, tenant isolation
rebase-realtime	The WebSocket engine: sync, broadcast channels, presence, table change broadcasts
rebase-storage	S3/GCS/local storage, uploads, TUS resumable uploads, image transformations
rebase-custom-functions	Custom API endpoints via file-based function discovery
rebase-cron-jobs	Scheduling recurring background tasks
rebase-webhooks	Outbound HTTP webhooks, HMAC signatures, retry and backoff
rebase-email	SMTP, templates, custom providers, the rebase.email singleton
rebase-entity-history	Entity versioning, change tracking, audit logs, reverting
rebase-admin	Navigating the admin panel, side drawers, URLs, embedding collection panels
rebase-ui-components	The @rebasepro/ui component library
rebase-design-language	The UI design language: tokens, color, typography, spacing, anti-patterns
rebase-studio	The Studio developer tools layer — SQL, RLS, storage, cron, schema visualizer, logs
rebase-cloud	Deploying and operating on Rebase Cloud — projects, managed databases, env vars, domains, logs, rollbacks
rebase-deployment	Self-hosting: Docker, Kubernetes, AWS, GCP, Azure, Hetzner, Railway and Render
rebase-local-env-setup	First-time setup: Node.js, pnpm, PostgreSQL, Docker

Two of these ask to be read unprompted. rebase-basics says it should be used whenever an assistant touches Rebase at all, and rebase-design-language says an agent must read it before creating or modifying any visual UI — that one exists because generated UI drifts from a design system faster than anything else in a codebase.

What a run looks like

  Found 21 Rebase skills

  ✓ Claude Code — 21 skills installed (+ 8 reference files) to .claude/skills

Skills ship from the @rebasepro/agent-skills package, which the CLI depends on, so the set you get matches your installed CLI version.