Plugins that connect AI coding agents to SafeGrd. SafeGrd backs up PostgreSQL, MySQL, MongoDB and SQLite databases, file trees and IMAP mailboxes. Each backup is encrypted with age before it leaves the host and locked with S3 Object Lock. A Fire Drill restores it on a schedule to prove it works.
/plugin marketplace add safegrd/agent-plugins
/plugin install safegrd@safegrd
Then run /mcp, choose the safegrd server and Authenticate. Your browser opens SafeGrd:
sign in, pick the organization Claude Code may reach, and approve. There is no token to
create or paste. The connection is listed under Tokens in the console, named after the
client, and you can revoke it there.
Once signed in, Claude Code can read your organizations, projects, surfaces, snapshots and Fire Drills. It can also ask for a backup or a drill now. It cannot delete, disable, re-route or re-key a backup: the server refuses those to an agent on every route. The remote server's tools are listed at safegrd.dev/docs/mcp.
| Skill | What it does |
|---|---|
backup-before-risky-change |
Before a migration, a schema change or a bulk delete, it takes a locked backup, runs a Fire Drill on it, and says whether it is safe to proceed |
check-backups |
Reports late backups, failed drills, surfaces never restored, and hosted storage use |
setup-safegrd |
Takes a host from nothing to a verified backup: the CLI, signing in, enrolling, the first backup and its restore test, and the local MCP server, plus the experimental guard hook when you ask for it |
The local server backs up, verifies and restores with that host's config and key. A restore writes only into a new or empty target.
claude mcp add safegrd-local -- safegrd mcpThe guard hook is experimental. It takes a locked snapshot before a command that can destroy data, such as
DROP TABLE or prisma migrate reset, and blocks the command if the snapshot fails. It
names the surface a project works on, so it goes in that project's .claude/settings.json,
not in this plugin. The setup-safegrd skill adds it only when you ask, and
safegrd.dev/docs/agents has the recipe for Claude Code,
Cursor and Codex.
codex/safegrd is the same plugin in the Agent Plugins format Codex and ChatGPT read:
plugin.json, mcp.json and the three skills, with the hook recipe for Codex. Like the
Claude Code plugin, it signs the user in to SafeGrd (OAuth, registered on first connection).
Add this repository as a marketplace and install it with /plugins; the marketplace file
is .agents/plugins/marketplace.json. The manifest follows OpenAI's plugin documentation
and has not yet been validated against a Codex release.
.github/workflows/check.yml validates the Claude Code marketplace and plugin with
claude plugin validate --strict. It also checks that every tool a skill names, in either
plugin, is served by the remote server, read from its public
server card. The check runs on
every change and daily.
MIT. See LICENSE. The SafeGrd CLI is in safegrd/cli.