SafeGrd CLI (safegrd)

safegrd backs up PostgreSQL, MySQL, MariaDB, MongoDB and SQLite databases, directory trees and IMAP mailboxes. Each backup is compressed with zstd and encrypted with age on the host that takes it, written to S3-compatible storage under Object Lock, and restored on a schedule to check that every table, row and file came back.

It is the CLI and daemon of SafeGrd. Documentation is at safegrd.dev/docs, and every command and flag is in the command reference.

• Backups go from your host straight to storage. The remote server receives a record of each backup (when it ran, its size, its digest and the date its lock ends) and the signed result of each restore test. The data never passes through it.
• You choose who holds the key. By default the remote server keeps your key sealed and releases it only to your enrolled hosts, so you can restore even after losing a host. With enroll --key-custody local, only you can decrypt these backups.
• Restores are tested. A Fire Drill restores the newest backup into a throwaway database or directory, compares its tables, rows and digests with what was backed up, and signs the result.
• An agent cannot remove a backup. safegrd guard takes a locked snapshot before a destructive command runs, and a personal access token cannot delete a backup or shorten its lock.

---

Installation

Install script (Linux and macOS)

curl -fsSL https://safegrd.dev/install.sh | sh

It detects your OS and architecture, downloads the matching release archive, checks it against the release's checksums.txt, and installs safegrd to /usr/local/bin (or ~/.local/bin when it cannot use sudo). The script is install.sh in this repository; https://raw.githubusercontent.com/safegrd/cli/main/install.sh serves the same file.

When it runs in a terminal, it then offers to connect the machine to a remote server: a browser login, a question about who holds the encryption key, then safegrd enroll. The login prints a URL and a code that you can open in a browser on any device, so it works the same on a server you reached over SSH or PuTTY. Without a terminal (CI, cloud-init, cron) it installs and stops.

Options are environment variables, and they go on the sh after the pipe. Written before curl they apply to curl, and the script never sees them:

curl -fsSL https://safegrd.dev/install.sh | VERSION=v0.0.3 sh                  # pin a release
curl -fsSL https://safegrd.dev/install.sh | SAFEGRD_INSTALL_DIR=~/.local/bin sh  # no sudo
curl -fsSL https://safegrd.dev/install.sh | SAFEGRD_NO_SETUP=1 sh               # install only

Variable	Effect
VERSION	Release tag to install. Default: the latest release.
SAFEGRD_INSTALL_DIR	Where the binary goes.
SAFEGRD_NO_SETUP=1	Install only; do not offer to log in and enroll.
SAFEGRD_FORCE_INSTALL=1	Download and install even when the release asked for is already installed.
SAFEGRD_PROJECT	Project ID or slug to enroll this machine into.
SAFEGRD_STORAGE	hosted or local: where backups go when the project has no bucket.
SAFEGRD_CLAIM	The claim code the console shows for this machine. It carries the project, where backups go and the surfaces to protect, so enrollment writes them into the config and asks nothing. Needs a release whose enroll has --claim; an older one stops and says so.
SAFEGRD_NODE_NAME	Name the machine is shown under.
SAFEGRD_KEY_CUSTODY	local keeps this host's key on the host. Without it the remote server keeps the key sealed and releases it only to your enrolled hosts.
SAFEGRD_SERVER_URL	Remote server to log in and enroll with. Default: https://safegrd.dev.

Homebrew (macOS and Linux)

brew install safegrd/tap/safegrd

The formula installs the same release archives as the script.

Go

go install github.com/safegrd/cli/cmd/safegrd@latest

Container image

ghcr.io/safegrd/cli runs on linux/amd64 and linux/arm64 and carries pg_dump, mysqldump and mongodump, and a PostgreSQL 18 server so a drill can restore a database into a throwaway cluster on the volume. The tag names the Postgres client: <version>-pg18 reads Postgres 18 and every older server. Releases are signed with cosign (keyless, from this repository's release workflow):

cosign verify ghcr.io/safegrd/cli:<version>-pg18 \
  --certificate-identity-regexp 'https://github.com/safegrd/cli/.github/workflows/release.yml@.*' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

packaging/compose runs the daemon beside a database, packaging/kubernetes runs daemon run --once as a CronJob, and packaging/helm/safegrd runs the daemon as a Deployment. Each names the release it was cut with, such as ghcr.io/safegrd/cli:0.0.10; change the tag to upgrade.

Release archives, or from source

Static binaries for Linux and macOS (amd64, arm64) and their SHA-256 checksums are on the releases page. To build:

git clone https://github.com/safegrd/cli.git && cd cli
make build

---

Quickstart

1. Create a config and a key

Skip this if you let the install script connect the machine: it already wrote the keypair and ~/.safegrd/config.yaml, and init refuses to overwrite them.

# writes an age X25519 keypair and ~/.safegrd/config.yaml
safegrd init \
  --database-url "postgres://postgres:password@localhost:5432/myapp_prod" \
  --storage local \
  --local-path ./backups \
  --retention-days 14

2. Connect to a remote server (optional)

# Prints a URL and a code: open it in a browser on any device, including over SSH
safegrd login

# Register this machine, generating its encryption key if it has none.
# --key-custody decides whether the server keeps a copy (safegrd) or not (local)
safegrd enroll --key-custody local

# A host that should encrypt but never decrypt: make the key pair on your laptop,
# then put only the printed public key in the host's encryption.public_key
safegrd keygen --out ~/.safegrd/keys/prod.key

# Headless / CI: a Personal Access Token instead of the browser
safegrd enroll --token "sg_pat_..."

# Or with the claim code the console's setup shows: the project, where backups go and
# the surfaces are written into ~/.safegrd/config.yaml, and the key stays on this machine
safegrd enroll --claim <code>
safegrd doctor                # every surface opens from this host; the console shows the result
safegrd daemon run --once

# Later, on an enrolled host: add the surfaces named for it in the console
# (Add surface on its row). Only adds; the old config is kept as config.yaml.bak
safegrd claim
safegrd daemon restart

# Verify the active session
safegrd whoami

3. Back up

# dumps, compresses, encrypts with age on this host, and writes to locked storage
safegrd backup

4. Run a Fire Drill

# restores into an empty database and checks every table's row count
safegrd verify \
  --snapshot snap-20260919-01 \
  --sandbox-target "postgres://postgres:password@localhost:5432/ephemeral_test_db"

5. Restore

safegrd restore \
  --snapshot snap-20260919-01 \
  --target "postgres://postgres:password@localhost:5432/myapp_recovered"

6. Protect several things on one host, unattended

Each entry under surfaces: in ~/.safegrd/config.yaml is one protected thing, and safegrd daemon install runs them on their schedules. A surface's credential block says where its secret comes from; the config never holds it:

surfaces:
  - id: app-primary
    type: postgres
    schedule: "@daily"
    credential:
      from: env                 # or: command (with run), file (with path),
      name: APP_DATABASE_URL    # or safegrd: the remote server holds it for this host
  - id: user-uploads
    type: files
    schedule: "6h"
    roots: ["/var/www/uploads"]

safegrd.example.yaml is the annotated version. The daemon reads its config when it starts, so run safegrd daemon restart after changing it.

7. Take a locked snapshot before a destructive command

safegrd guard backs up one surface, waits until the snapshot is uploaded and locked, and then runs the command. If the backup fails or the snapshot is not locked, the command does not run and guard exits 3.

safegrd guard --surface app-primary -- psql "$APP_DATABASE_URL" -c 'DROP TABLE sessions'
safegrd guard -- terraform destroy

# The commands a hook treats as destructive, and a check that takes no backup
safegrd guard --list
safegrd guard --matches "npx prisma migrate reset"   # exit 0: it matches

Storage on the host's own disk, or a bucket with worm_mode: NONE, cannot lock a snapshot, so guard refuses there unless you pass --allow-unlocked.

A locked snapshot of the surface younger than --max-age stands in for a new backup, so an agent running several destructive commands in a row backs up once. --check-only takes no backup at all and refuses the command unless such a snapshot exists:

safegrd guard --max-age 30m -- psql "$APP_DATABASE_URL" -c 'TRUNCATE sessions'
safegrd guard --max-age 1h --check-only -- terraform destroy

safegrd list --json prints every snapshot with its lock date, for scripts.

8. Back up a directory tree, and get one file back

File backups are incremental: the first run of each month uploads every file, and every run after it uploads only the chunks that changed. Every kept version of every file is searchable, and one file restores on its own.

safegrd backup --files /srv/app --exclude '*.tmp,node_modules/*'

# every kept version of a file, numbered from the oldest
safegrd find etc/nginx/nginx.conf

# one version of one file, or a directory as it was in one snapshot
safegrd restore --path etc/nginx/nginx.conf --version 2 --target-dir ./out
safegrd restore --snapshot snap-20260919-01 --path 'var/www/**' --target-dir ./out

# every pack present and every listing consistent, without restoring anything
safegrd check --all
# a bucket over the S3 API (Supabase Storage, or any S3-compatible store)
safegrd backup --files s3://avatars --source-endpoint https://PROJECT.storage.supabase.co/storage/v1/s3 \
  --source-region eu-central-1 --source-access-key env:S3_KEY --source-secret-key env:S3_SECRET
# and back into the bucket, each object with the content type it had
safegrd restore --snapshot snap-20261012-01 --target s3://avatars --source-endpoint https://PROJECT.storage.supabase.co/storage/v1/s3 \
  --source-region eu-central-1 --source-access-key env:S3_KEY --source-secret-key env:S3_SECRET

--format tar keeps one archive per backup instead. What is locked, for how long, and what a restore brings back: safegrd.dev/docs/surfaces/files.

---

AI agents (MCP)

safegrd has two MCP servers. Neither has a tool that deletes a backup, changes where backups go, or takes a private key.

The local server, safegrd mcp, runs on a host over stdio and uses that host's config and key. Its tools are status, list, doctor, backup, verify, restore and export, each running the command of the same name. restore writes only into a new or empty directory or an empty database.

claude mcp add safegrd-local -- safegrd mcp

# a host that runs the container image: mount the volume it enrolled with
claude mcp add safegrd-local -- docker run -i --rm \
  -v safegrd:/home/safegrd/.safegrd ghcr.io/safegrd/cli mcp

The remote server, https://safegrd.dev/mcp, reads every host's backups and Fire Drills and asks for a backup or a drill now. It takes a personal access token from the console's Tokens page as Authorization: Bearer sg_pat_..., and is included on the paid plans.

claude mcp add --transport http safegrd https://safegrd.dev/mcp \
  --header "Authorization: Bearer sg_pat_..."

The Claude Code plugin in safegrd/agent-plugins adds the remote server with skills that back up and run a Fire Drill before a risky change:

/plugin marketplace add safegrd/agent-plugins
/plugin install safegrd@safegrd

Both servers are listed in the MCP Registry as dev.safegrd/safegrd, and the remote one on Smithery. Cursor and VS Code configs, and every tool's arguments: safegrd.dev/docs/mcp.

---

What it backs up

• PostgreSQL, incrementally: the schema comes from pg_dump and the rows stream over binary COPY from the same snapshot, so arrays, enums, foreign keys, views, triggers and sequence positions all come back. Each run uploads only the chunks of each table that changed, restore --table loads chosen tables into a running database, and --change-log skips reading tables nothing wrote since the last run. --format tar keeps one archive per backup. Needs a pg_dump at least as new as the server on the host.
• SQLite, incrementally: copied through SQLite's online backup in one read transaction, so transactions still in the WAL are in it and writers carry on. The copy keeps the database's pages in place, so each run uploads only the chunks whose pages changed. --format tar keeps one archive per backup. Needs nothing installed on the host.
• MySQL and MariaDB through mysqldump --single-transaction, and MongoDB through mongodump --archive, one archive per backup.
• Directory trees, incrementally: each run uploads only the chunks that changed, every object is locked once when it is written, and safegrd find lists every kept version of a file for restore --path --version.
• IMAP mailboxes, every message as RFC 5322 mail, checked by SHA-256 on every drill.
• Locked storage: S3 Object Lock in compliance mode, or governance mode where a lock may need lifting. Storage on the host's own disk, and a bucket with worm_mode: NONE, keep backups without a lock.
• Fire Drills: restore backups on a schedule, count what came back, and sign a record of each test.

---

Development

Go 1.25 or newer. make with no target lists everything.

make build      # compile bin/safegrd
make test       # unit and integration tests, race detector
make ci         # the exact gate CI runs: gofmt check, tidy check, vet, race tests
make dist       # cross-compile release archives + checksums into dist/
make fmt tidy   # the developer-facing fixers (these rewrite files; `ci` never does)

CI runs make ci on every push and pull request, so a green local run means a green pipeline. Tagging v* runs make dist VERSION=<tag without v> and publishes the archives and checksums.txt as a GitHub Release.

---

License

Business Source License 1.1, with an Additional Use Grant. See LICENSE.

In short: read, modify and build it freely, and use it for evaluation anywhere. Production use is included with a SafeGrd account on any plan, the free one too. Listing, verifying, restoring and exporting your backups is always permitted, with or without an account. Each version becomes GPL-3.0-or-later four years after it is published. For a commercial licence, contact support@safegrd.dev.