Stop retyping the prompt that works.

A place to keep the prompts and skills you use with AI tools. On your own, you save a prompt once and type //name in ChatGPT, Claude, Gemini or Grok to use it. As an organisation, you keep them in one place, review each version before anyone runs it, and install it into every coding agent with one command.

A skill is a folder of instructions, led by a SKILL.md file, that a coding agent such as Claude Code or Antigravity loads when a task matches its description. Cursor reads the same instructions as a rule. A skill tells the agent how your team does one job, like a deploy or a migration. Deftro keeps every version of it.

A prompt is text you send to a chatbot yourself, like ChatGPT, Claude or Gemini, usually with a few blanks to fill. A skill is a set of instructions a coding agent reads on its own when the task calls for it. Deftro keeps both in one library, with the same versions and the same review.

No. If you use a chatbot in a browser, you save a prompt on Deftro, add the browser extension, and type // in the message box. There is no terminal and nothing to configure. The command line and coding-agent installs exist for engineering teams and stay out of your way.

Yes for one person, and for a team of up to ten members. A company pays when it wants the review gate, health checks and an exportable Activity record.

In Postgres, one row per file keyed by its SHA-256. Everything is private by default. Bodies are never sent to a model provider.

Add the browser extension, open ChatGPT, Claude, Gemini or Grok, and type // in the message box followed by the start of the prompt's name. Your saved prompts appear above the box. Pick one with Enter.

No. It puts the prompt in the message box and stops. If the prompt has blanks like [TOPIC], a short form asks for them first. The text waits in the box until you press send yourself.

A published version waits until an admin or project lead reviews it. They see a diff against the last approved version and the health check notes, then approve or request changes. Every decision is a row in Activity.

It cannot. Approved versions are immutable, enforced by a database trigger. A change is a new version with its own diff, reviewer and record. Installs stay pinned to the version they were given.

The health check flags the careless cases: a leaked key, an obvious "ignore previous instructions", hidden text. It does not catch a determined attacker, and we never say it does. The product is the review, the diff and the record.

Not by default. A skill runs with your agent's access to your files, terminal and credentials, so read it before you install it. Deftro shows each version's health check notes, the diff against the last approved version and who reviewed it, so your team decides what runs. We never certify a skill as harmless.

Run deftro install <handle>/<slug> in your project. The bv CLI writes the file each agent there reads: .claude/skills/<slug>/SKILL.md for Claude Code, .cursor/rules/<slug>.mdc for Cursor and .agents/skills/<slug>/SKILL.md for Antigravity. Add --agent claude to pick one. Each install is pinned to the approved version and recorded in bv.lock.

The bv CLI writes Claude Code, Cursor and Antigravity formats directly. Any MCP client can use @deftro/mcp, and the browser extension inserts approved artifacts into ChatGPT, Claude, Gemini and Grok with a //alias.

Yes. Create a scoped token and paste the agent brief into Claude Code, Cursor or Codex. The agent can then search, install, publish and sync skills, and can never approve, delete or make anything public.