Company / Changelog
Changelog
What has shipped, newest first. CLI releases are at github.com/safegrd/cli/releases.
31 October 2026: from a connected host to a first backup
New
safegrd restore --table <schema.table> --rows "<condition>"puts the rows a condition matches back into a live PostgreSQL table by primary key: a deleted row comes back, a row the target still has is left as it is unless you pass--overwrite, and the table's other rows and the database's other tables are untouched.verify --queryreads rows out of a backup as CSV first, if you want to look before you write. Needs the next CLI release.- You can restore rows from the console too: the Restore dialog of a PostgreSQL backup takes a table and a condition, and the Tables tab has Restore rows on each table. Your host or a machine SafeGrd starts runs it, as for any restore from the console. Needs the next CLI release on the host.
- You can restore one table of a MySQL, MariaDB or SQLite backup, or one collection of a MongoDB backup, into a database that has other tables:
safegrd restore --snapshot <id> --table orders --target mysql://…. The table is created as the backup defined it, with its indexes and triggers, or loaded into an empty one of the same columns. From the console, the Tables tab of each of these surfaces has Restore table, and the Restore dialog takes chosen tables. Works on incremental backups (the default); a full backup restores whole. Needs the next CLI release. - Each surface's panel shows its trend over the last 30 days under Trend: how much the backup grows a day, the day your hosted storage fills at that rate, the last run's time and speed against the median of the last 30 runs, the tables that drove an incremental backup's uploads, and what a run noticed against the backup before (a mail folder the server rebuilt, directories gone, an extension gone from the top of the list). From the figures each backup already reports; nothing is read out of a backup.
- You can pin a required Google sign-in to your Workspace domain under Settings → Security policy: a sign-in from another domain or a consumer account is refused, and so is an invitation to an address outside it. Members of an organization that requires a provider now get a session of 1 day instead of 7, so an account the provider suspends is out within a day.
safegrd restore --table public.orders --as public.orders_restoredputs a PostgreSQL table's backup beside the live one, under the name you give, so you can compare the two with SQL and copy back what you need. Needs the next CLI release.safegrd listdescribes each snapshot from SafeGrd's record of it, read in one request, instead of downloading each one's metadata file: 40 snapshots on hosted storage listed in 40 seconds before.list --project <id>lists one project's snapshots. Needs the next CLI release.- A Heroku sign-in SafeGrd holds for a surface lists the account's databases again once a day, so a database added in Heroku since the sign-in shows under Not backed up without connecting again.
- The Hosts tab says which databases a Heroku, DigitalOcean or Supabase sign-in listed have no surface, with the date of the sign-in, and offers to back each up or mark it Not needed. A sign-in lists the account as it was then; connecting again refreshes the list.
- Alerts go to Microsoft Teams and PagerDuty. A Teams workflow URL pasted under Settings → Alerts gets each alert as a card. A PagerDuty integration key opens an incident for a failure or a warning, one per surface and condition, and the recovery alert resolves it; the test alert opens one and resolves it at once. Each alert now carries its
severity, and a recovery names the condition itclears. - The setup checklist shows your next step until your first backup lands. For a host with no surface yet, that is Add surface, then the commands that take its first backup. A surface connected is ticked once you add a surface. A host alone no longer ticks it.
- You can see which hosts enrolled and never started backing up. A host enrolled more than 24 hours ago with no surface, no daemon and no backup shows Not backing up on the Hosts tab, with what to run. You get one alert for it. A host enrolled only to restore needs nothing more.
- After
safegrd enrollyou get one set of next steps, however the host enrolled, and they lead to the daemon:safegrd doctor,safegrd daemon run --once, thensudo safegrd daemon install. The install script adds no steps after them. Needs the next CLI release. safegrd tryno longer starts with a warning when you type a URL with its password. It keeps the URL in a file beside the config that only you can read, so the config holds no password. Needs the next CLI release.- Sign-up asks how you found SafeGrd, and you can skip the question.
/signup?plan=freesays what Free is: no card, 14 days of Growth first, then Free for as long as you want.
Fixed
- SafeGrd now releases a key, a credential or a bucket download link only after recording the release in your key access log.
- A backup or a drill on hosted storage failed when the storage stayed busy for more than a few seconds. It now waits longer before it gives up. Needs the next CLI release.
- A check of another tool's backups run with
safegrd backupno longer asks hosted storage about an upload first, so it is neither warned about the storage quota nor refused at it. Needs the next CLI release.
30 October 2026: the Evidence Pack as a PDF
New
- You can download the Evidence Pack as a PDF for an auditor who reads on paper, from the Fire Drills tab or the pack's own page, beside the JSON. It holds the summary for a questionnaire, the controls and the evidence for each, recovery point and restore time, key custody, every drill record, Threat Shield events and every snapshot's lock. The PDF has no signature of its own, so verify the JSON file (how to verify, a sample).
- You can publish a public page for a project that lists each of its surfaces with its kind, last backup and newest drill. Publish it from Project pages on the sharing card of the Fire Drills tab. SafeGrd signs the page, and the browser checks it like a certificate. It shows how many surfaces the project has, and their names only if you choose to show them (details).
- You can have SafeGrd check the backups another tool makes, with no host of yours. Under Back up on SafeGrd, tick Check the backups another tool makes and give the bucket, the prefix and a key that can list and read it. On the surface's schedule, SafeGrd records the newest dump as a backup made by that tool, and its drills load the dump into a sandbox. SafeGrd holds the key, encrypted, and uses it only on the machines it starts for these checks and drills. A drill's machine gets that key and never a database credential, and the key access log shows each use (details).
- You can have SafeGrd check Coolify's MySQL, MariaDB and MongoDB backups where Coolify puts them, as it already did PostgreSQL's. Set
watch:on a surface:type: mysqlreads plainmysqldumpfiles andtype: mongodbgzippedmongodumparchives. Each dump counts once and is drilled into a sandbox of its engine. A dump cut short in the upload is a failed backup, and a sandbox of the wrong engine is refused before anything connects. Needs the next CLI release (MySQL, MongoDB). - You can keep daily, weekly and monthly copies of a one-off backup, as a surface does.
safegrd backup --database-url,--filesor--emailnow keeps the first backup of each day, week and month longer. Set the new--keep-daily,--keep-weeklyand--keep-monthlyflags, or the config'sdefaults.keep_daily,keep_weeklyandkeep_monthly. The run says which copy it is ("Kept as the weekly copy: locked until 2026-11-27"). A run with--retention-daysand no--keep-*flag is locked for exactly those days. Needs the next CLI release (how it decides). - You can see in a database's drill record which tables changed since the backup before. It lists each table whose row count moved, largest move first, the tables added and gone, and how many did not change. The drill's detail in the console shows it, and the list is signed with the record. The verifier page checks such a record now. Checking it with
safegrd verify historyon your own machine needs the next CLI release (the record).
Fixed
- A drill certificate for another tool's backup no longer says "Decrypted and read in memory", since that backup is not encrypted.
- The Evidence Pack page no longer shows every surface's recovery point as "0 h old" and its restore time as "not drilled".
29 October 2026: an auditor role
New
- You can give your auditor their own sign-in: invite them as an Auditor from Settings, Team members. They see the drill records, each drill's signed certificate and the Evidence Pack, and nothing else: no surfaces, hosts, snapshots, storage, members or settings, and nothing that starts or changes anything. With a token of their own, they can run
safegrd verify historyto verify the chain on their own machine (how it works).
27 October 2026: a copy to your bucket on the daemon's schedule
New
- You can have the daemon copy every snapshot, still encrypted, to a bucket or directory you own, so a copy of your hosted backups needs no crontab line. Set
daemon.export_toin the host's config, and the daemon runssafegrd exportafter its backups, daily unless you set a schedule. The Hosts tab shows when the last copy was made, and a run that failed. Needs the next CLI release (the keys, the runbook).
Fixed
- Back up now on a surface that watches another tool's dumps no longer stays Requested for a day when the check finds no new dump. Needs the next CLI release.
- After a failed backup from
safegrd daemon run --once, the console no longer shows the surface as Running for a day or refuses Back up now on SafeGrd. Needs the next CLI release.
25 October 2026: a first backup before you sign up
New
- You can run a backup or a drill on SafeGrd at once, whatever normally runs the surface, for a host that is down or an audit tomorrow. Press Back up now on SafeGrd or Drill now on SafeGrd on the surface's Details, or pass
on: safegrdto the MCP tools. Each paid plan includes a number of these runs a month, shared by the organization, and the console and the pricing table say how many. A drill your plan's cadence already allows does not count (details). - You can see how many minutes a month a surface's schedule will use on SafeGrd when you save its connection string, beside the egress estimate. Each plan says how many surfaces SafeGrd backs up from a credential it holds. The limit is checked when you add one, never afterwards.
- You can take a real backup before you have an account:
safegrd try postgres://…, or a MySQL, MongoDB or SQLite URL, or a directory. It is encrypted on your host, uploaded to SafeGrd's storage and restored in memory from there to check it, and it prints the tables and rows it read back. The backup is kept for 24 hours.safegrd signupopens the sign-up page: sign up to keep the backup and see its drill in the console. The host, its backup and its daily schedule are yours once the account exists. Needs CLIv0.0.59(details).
Fixed
- Signing up with Google or GitHub from the page
safegrd signupopens now keeps your trial backup, as a password sign-up does. - Opening that page signed in to an account that has an organization now shows the refusal with a link to your console, without sign-up buttons.
- A host where you ran
safegrd trycould not join an organization you already had. Runsafegrd login, thensafegrd enroll, on that host: it enrolls as a new host in your organization, with the same surfaces and the same key. A SafeGrd-managed key stays SafeGrd-managed, and the config records the mode inencryption.custody. Needs the next CLI release.
24 October 2026: a Supabase project's Storage beside its database
New
- A Supabase project you add with the sign-in now shows under Supabase project in the surfaces list. Its Storage shows beside it once you add that as a bucket with an S3 key, as the dialog shows (details).
- You can back up every bucket an S3 key lists in one run:
safegrd backup --files s3://*.safegrd restore --target s3://*puts each file back in its bucket. Both take--source-session-tokenfor a temporary key. Needs CLIv0.0.59.
23 October 2026: WordPress in your own bucket without Object Lock, and large databases on strict hosts
New
- You can back up a WordPress site whose host stops a request before the database is read. If the host stops the plugin's first attempt, the plugin reads the tables over several requests, so they are read at different moments. The plugin's log, Tools, SafeGrd and the snapshot in the console say so. For one consistent snapshot of a large site, run
wp safegrd backupfrom the host's cron (details). Needs the next plugin release. - You can back up a WordPress site to your project's own bucket without Object Lock (
worm_mode: NONE), such as DigitalOcean Spaces. Set the bucket up in the console with a key SafeGrd holds. The site backs up there once Test passes. These backups are not locked: the plugin,wp safegrdand the console show them as Not locked, and they stay until you or the bucket's lifecycle rules delete them. SafeGrd's drills read them there, and the bucket needs no versioning (details). Needs the next plugin release.
22 October 2026: snapshot notes in the console
New
- You can back up to a directory on an SSH server over SFTP (
storage.type: sftp), for every surface type and both formats, with restores, Fire Drills andsafegrd prune. An SFTP server has no Object Lock, so the storage takesworm_mode: NONEonly, and each backup is kept until the date it records. The CLI checks the server's host key againstknown_hostsor a pinned fingerprint before it sends anything (details). Needs CLIv0.0.58. - You can see a snapshot's notes, such as
release=v2.3, in the console, on the snapshot and on each drill of it. A published certificate does not show them. - If buckets you set in the console replace
copies:in a host's config, the surface's panel names the copies that are not written.
Fixed
- A drill check the sandbox could not run now shows as not run, with the reason, and is counted apart from the failures.
- When one of your drill checks fails and a data check cannot run, the drill alert now names the failed check instead of saying the snapshot did not restore.
- A WordPress backup that failed partway now says the run failed and how long the part it stored is locked, where the console said "Nothing stored".
21 October 2026: a WordPress site or a database SafeGrd backs up, in your own bucket
New
- You can back up a WordPress site to its project's own bucket, on every plan. Set the bucket up in the console with a key SafeGrd holds. The site backs up there once Test passes. The site never holds the bucket's key, and SafeGrd's drills read the backups from your bucket. A bucket without Object Lock keeps the site on hosted storage, and the plugin says why (details).
- You can add a PostgreSQL database SafeGrd backs up to a project that stores in its own bucket, under the same conditions. Its incremental backups go to your bucket, and SafeGrd's drills restore them from there. Before, the console refused it (details).
- You can see where a WordPress site's or a database's next backup goes on its panel: your bucket by name, or hosted storage with the reason.
Fixed
- A backup through a transaction pooler (Supabase's port 6543, or PgBouncer in transaction mode) could finish with no tables, or leave tables out, and still report success. It is now refused, and the error says to use the session pooler or connect directly. Needs CLI
v0.0.58. - An incremental backup to a local directory now stops before it fills the disk, as a full backup does. Needs CLI
v0.0.58. - A base backup no longer fails when one page of a cluster with data checksums fails its checksum. It keeps the base and names the file and how many pages failed. A logical backup that reads such a page names the table and the block. Needs CLI
v0.0.58. - A base backup of PostgreSQL 13 or 14 no longer fails with "unexpected EOF". Needs CLI
v0.0.58. - One damaged file in a WAL archive no longer makes part of its WAL unreachable, and
wal status --checkreports any damage it finds. This release does not read WAL archives written by earlier releases. Needs CLIv0.0.58. - A cluster restored from a base backup taken on a standby now promotes, instead of staying a standby that streams from the live primary. Needs CLI
v0.0.58. - A cluster restored by
safegrd wal restorerefuses to start if itsbackup_labelis deleted, instead of starting from the wrong checkpoint. Needs CLIv0.0.58. - The recovery document beside a copy in a second bucket now names the bucket it is in and that bucket's lock, where it named your own bucket. Needs CLI
v0.0.58. - A backup that
safegrd copies syncwrites to a second bucket now gets a recovery document and a manifest that name that bucket and its lock. Needs CLIv0.0.58. - An email backup no longer holds the mailbox in memory, which took about 730 MB for a 500 MB mailbox. It needs free disk about the mailbox's size in the temporary directory instead. Set
TMPDIRto use another disk. Needs CLIv0.0.58. safegrd prunenow prunes incremental backups in a bucket without Object Lock (worm_mode: NONE), where every object failed with "could not read its lock". Needs CLIv0.0.58.
Changed
- Your incremental PostgreSQL backups read several tables at once, all at the same moment of the database. Set
parallelon the surface to change how many. A full backup (one archive) still reads one table at a time (details). Needs CLIv0.0.58. - Your incremental MySQL or MariaDB backups can read several tables at once, all at the same moment of the database. Grant the account
FLUSH_TABLES(MySQL 8.0.23 and newer) orRELOAD, so it can hold writes for that moment. Writes wait while the backup starts, usually under a second, and the backup prints how long. Without the grant it reads one table at a time and says so (details). Needs CLIv0.0.58. - Your incremental MongoDB backups read several collections at once. Set
parallelon the surface to change how many. Needs CLIv0.0.58. - You can restore an incremental MySQL or MariaDB backup over several connections at once. Set
restore --parallelorparallelon the surface, and1loads one table at a time. Sandbox drills load the same way. Needs CLIv0.0.58. - You can restore an incremental PostgreSQL backup over several connections at once. Set
restore --parallelorparallelon the surface, and1keeps the restore in one transaction. Sandbox drills load the same way, so a drill takes as long as a restore (details). Needs CLIv0.0.58. - You can put notes on a snapshot, such as
release=v2.3:backup --noteas it is taken, orsafegrd notelater. SafeGrd keeps them outside the locked backup, so they can change, andsafegrd listshows them (details). Needs CLIv0.0.58. - You can set how hard a surface's backups are compressed, from
zstd:1tozstd:19, withcompressionin the config orbackup --compression. A restore needs no setting (details). Needs CLIv0.0.58. - You can restore to a WAL position or a transaction:
safegrd wal restore --to-lsnor--to-xid.--exclusivestops just before it, and--timelinefollows the timeline you choose (details). Needs CLIv0.0.58. - A sandbox Fire Drill now has the database check the data it restored:
amcheckon PostgreSQL,CHECK TABLEon MySQL,validateon MongoDB. A check the sandbox cannot run is shown as not run, with the reason, and does not fail the drill (details). Needs CLIv0.0.58. - You can have PostgreSQL's
archive_commandreturn as soon as a WAL file is on the host's disk: setwal.spool, and the daemon uploads the spool.wal.queue_maxcaps the WAL waiting on the host, so an unreachable archive cannot fillpg_waland stop the database. WAL let go past the cap is listed bywal statusand alerted (details). Needs CLIv0.0.58. - Your incremental backups are now written to every bucket during the run, so your provider no longer bills a copy from your own bucket as egress. A bucket that missed part of the month is still filled from your own bucket (details). Needs CLI
v0.0.58. safegrd prunekeeps the newest backup whose drill passed when the oldest one it would keep failed its drill. Needs CLIv0.0.58.- You can prune your other buckets too:
safegrd pruneprunes each after your own, by that bucket's locks andmin_keep. Setexpire_after_lockon a copy for the daemon to prune it daily.findandprunetake--from-bucket. - If buckets you set in the console replace
copies:in a host's config,safegrd doctorsays which copies are not written, and the host reports them to SafeGrd.
20 October 2026: restore from the console, and a second bucket
New
- You can restore a backup from the console on every paid plan: press Restore on a snapshot, choose what to restore, and type the target database's URL. Your host runs it at its next check-in, and the new Restores tab shows how it went. On Free, the button shows the command to run on your host (details). Needs CLI
v0.0.57on the host. - You can have every backup written to up to two more buckets, each with its own access key, lock and retention. Add them under More buckets on the project's Storage card on Growth and Scale, or under
copies:in the host's config on any plan (details). Needs CLIv0.0.57. - You can fail a drill whose newest row is older than you expect: a drill check's query can name
'${SAFEGRD_SNAPSHOT_CREATED_AT}', the time the backup was taken. Before, only a command check in the host's config had it, so drills on SafeGrd's machines could not check this. From a template in the Drill checks dialog writes this check, and two other common ones, in your database's SQL (details). Needs CLIv0.0.57on the host. SafeGrd's machines already run it. - You can see what each of your buckets lacks with
safegrd copies status, and fill it in withsafegrd copies sync.restore,verifyandlistread a copy with--from-bucket, anddrill.source: rotatedrills each bucket in turn.
Fixed
safegrd restore --schemanow prints the chosen schemas' table and row counts, where it printed the whole backup's.
19 October 2026: pause a surface's schedule
Fixed
- A Supabase database added through the Supabase sign-in no longer fails every backup with "row-level security hides rows", and its backups now include the
authandstorageschemas. Needs CLIv0.0.55, which SafeGrd's machines already run.
New
- An agent connected to SafeGrd can now help you add a database. Ask it to back one up, and it gives you a console link that opens Back up on SafeGrd on the right project, with Supabase, DigitalOcean or Heroku sign-in ready. For a database on your own machine it gives you the commands. You finish in the console, so no password passes through the chat (details).
- You can connect Claude, ChatGPT or another agent to SafeGrd on the free plan. The remote MCP server and its sign-in were paid-only. What an agent can ask for follows your plan, as it does in the console (details).
- You can pause a surface's schedule from Edit in the console, with a reason and a time to resume on if you want one. While it is paused, no scheduled backup or drill runs and the surface is not reported Overdue. Back up now and Drill now still run. Resume on the row ends it (details). A host's surface needs CLI
v0.0.54. A surface SafeGrd backs up pauses now. - Days a surface was paused show as paused in its 30-day history, not as missed backups. The worst gap leaves paused time out and says how much. The recovery figures and the Evidence Pack list each pause with who paused it and why (details).
18 October 2026: alerts and an activity log in the console, and daemon metrics
New
- Sign in to the remote MCP server from claude.ai, Claude Desktop, Claude Code and ChatGPT: add
https://safegrd.dev/mcpby URL, and approve in the console for one organization. No token to paste. The sign-in is listed under Tokens, where you revoke it (details). - You can update a host's CLI from the console: an owner or admin clicks Update on the host's row. The host checks the release against its checksums and installs it between backups, with no restart. If the new release does not check in, the host goes back to the one it ran before, and Go back does the same on request. A host installed earlier needs
install.shrun once (details). Needs CLIv0.0.53. - You can have SafeGrd check the PostgreSQL backups another tool makes, such as Coolify's, where that tool puts them. A surface with
watch:counts each new backup's tables, goes Overdue when the tool stops writing, and drills the newest into a sandbox. The certificate names the tool (details, guide). Needs CLIv0.0.53. - A guide: Supabase backup tools compared, eleven ways to back up a Supabase project, with which keep the Storage files and which test the restore.
- A guide: restore a Postgres backup without bringing back deleted users or starting its pg_cron jobs, with
pg_restoreor SafeGrd. - You can restore an older backup without bringing back rows erased since. Name the tables under
erasures:, and the restore deletes those rows again before it commits. The record is kept encrypted beside your backups, a restore that cannot read it stops, and a drill checks it can be read (details). Needs CLIv0.0.53. - You get your pg_cron jobs back after a PostgreSQL restore, turned off, each printed with how to turn it on.
--keep-cron-activekeeps them as they were. Before, a backup left them out, and a restore lost every schedule without saying so (details). Needs CLIv0.0.53. - You can fail the drill of a database that had stopped changing. A command drill check gets
SAFEGRD_SNAPSHOT_CREATED_AT, when the backup was taken, to compare with your newest row (details). Needs CLIv0.0.53. - Your drill checks now run on a SQLite surface with no
drill.sandbox_url, on a plan with sandbox drills: the daemon drills it into a new file in its state directory. Before, it drilled in memory and skipped the checks (details). - You can keep your own copy of a host's SafeGrd-managed key.
safegrd keygen --held --out FILEon an enrolled host writes it, or with--snapshotthe key one snapshot was sealed to. SafeGrd keeps the key sealed as before, and your key access log lists the copy (details). Needs CLIv0.0.51. - Supabase backup compares the cost of Supabase's point-in-time recovery add-on with a copy kept outside the project, from Supabase's own prices read on 18 October.
- You can scrape the daemon with Prometheus: set
daemon.metrics_addr, and it serves each surface's last backup, failures, next run and last drill (details). Needs CLIv0.0.51. - You can have a Healthchecks.io or Cronitor check alert you when backups fail or the daemon stops. Set
ping_urlon a surface to ping it before and after each backup, anddaemon.ping_urlto ping after each pass. Needs CLIv0.0.51. - You can see who changed what, and when, in the Activity log under Settings: sign-ins, members, access tokens, hosts, projects, storage and security settings. Owners and admins see it, and can export it, and every key release, as CSV or JSON (details).
- You can restore what is still good from an incremental files snapshot with damaged or missing parts.
safegrd restore --salvagerestores every file that still verifies and lists the ones it left out. Needs CLIv0.0.51. - A comparison: SafeGrd vs Barman. The pgBackRest and Databasus comparisons list point-in-time recovery for self-hosted PostgreSQL and SQLite.
- If your Supabase project is down and cannot restore itself, you can move to a new project from your own backup, with the files in Storage and the settings a dump leaves out (guide).
- Guides: database backups an AI coding agent cannot delete, and restore-test evidence for SOC 2 and security questionnaires.
- Where to run it has the steps for Docker Compose, a Kubernetes CronJob, the Helm chart and GitHub Actions.
- You can see how far a backup read from a standby lagged the primary. The CLI's output and the snapshot's panel say so, with how long before the backup its newest replayed transaction committed on the primary (details). Needs CLI
v0.0.51. - You get a
duplicate_hostalert naming both machines when two machines report as one host, as two copies of one config do (details). Needs CLIv0.0.51. safegrd doctornow warns you when a MySQL or MongoDB surface's account can write, as it already did for PostgreSQL, and prints the account to create instead. The MySQL and MongoDB pages list its grants. Needs CLIv0.0.51.- You can give an agent the docs as Markdown: add
.mdto any docs page's address, or read /llms-full.txt for all of them. - You can search the docs from the box at the top of the sidebar. It finds the sections of every docs page that hold each word you type.
- You can see every alert from the last 30 days under Alerts in the console's top bar, newest first, with the surface it is about. It lists them whether or not you set up a webhook or email, and every member sees it, viewers included (details).
Changed
- Two MCP tools are renamed, so every tool is a verb:
get_storage_usageandget_drill_stats. - You get one alert when a surface starts failing, another if its error changes, and one when it backs up again, instead of one for each failed run. A surface the daemon runs no longer sends two alerts for its first failure (details).
- The activity log has its own page, Activity, in the console's sidebar, for owners and admins. A host reading its own hosted storage is listed in the key access log only.
- A running daemon says once in its log when its binary was replaced by an upgrade, with the command that runs the new one. Needs CLI
v0.0.51. - You no longer restart the daemon after adding a surface: it reads its config again when the file changes, whether you ran
safegrd claimor edited it by hand. Needs CLIv0.0.51. - You can see on a surface's panel what it keeps in storage now and once its copies have rolled over. It also shows how long each backup is held and, for your own bucket, about what that costs a month at the provider's list price.
- A project's storage is set under Projects: each project's Storage button opens it below the table. The separate Storage tab is gone, and links to it open there.
- Your key access log keeps each release for 90 days, with the address and client that asked, then deletes it, as the privacy notice says.
- The activity log now also records sign-ins refused for a wrong password or two-factor code, surfaces added on SafeGrd and retired, and a surface's connection changed or removed. A retired surface was listed as a host decommissioned.
- You can see how many backups a month a schedule starts before you save it, in the schedule field of a surface's settings.
- The setting that lets your hosts delete backups whose lock has ended says what it allows: any backup whose lock has ended, never a locked one (details).
- An organization deleted 12 months ago is now erased, with its backup records, as the terms say. Its backups in your own bucket are not touched.
Fixed
- A Fire Drill of a PostgreSQL database that uses pg_cron no longer fails at
CREATE EXTENSIONand reports the backup as broken. The drill holds every pg_cron job and sends no pg_net request. On hosts this needs CLIv0.0.53. - A restore of a whole PostgreSQL database now brings back the privileges on the database itself. Before, a database that let only its own roles connect came back open to every role on the server.
--no-ownerstill restores no privileges. Needs CLIv0.0.51. safegrd restore --tableof a table in a schema the target database lacks no longer fails withschema does not exist: it creates the schema with the table. It also takes a name quoted aspsqlprints it, such as"Sales Team"."Order Items"(details). Needs CLIv0.0.51.- A
format: tarfiles backup whose root is a symlink to a directory (a deploy'scurrent, say) now backs up the directory the link points at, where it said "Backed up 0 files". Needs CLIv0.0.51. - A
format: tarfiles backup of a directory holding the host's local backup storage now leaves that storage out and says so, as the incremental format already did. Needs CLIv0.0.51. - Two surfaces with the same id in a host's config passed
safegrd config validate, and the daemon backed both up under that id, mixing their backups. Validation now fails naming the id, and the daemon backs up neither until each has its own. Needs CLIv0.0.51. - A PostgreSQL restore now brings back settings made on the database itself (
ALTER DATABASE ... SET search_path, atimezone) and a role's settings in that database (ALTER ROLE ... IN DATABASE ... SET). Before, your application met the server's defaults. The restore names any setting the restoring role may not apply. Needs CLIv0.0.51. - A sandbox Fire Drill of a PostgreSQL database that uses table inheritance no longer fails when the restore was right. Needs CLI
v0.0.51. - A PostgreSQL backup left out the contents of large objects (
pg_largeobject) without a warning, and the Fire Drill still passed. Large objects are now in the backup in both formats, and a restore brings each back under its own oid with its owner. Both drills check their bytes, and--to-sqlwrites them as SQL. A large object the backup role may not read is named in a warning, and the drill fails on it (details). Needs CLIv0.0.52. safegrd verify history --fileand the verifier page now check an Evidence Pack with more than one host, or one for a quarter, as the pack says to (details). The CLI half needs CLIv0.0.51.- A mailbox with a message whose headers do not parse, as old spam often has, now passes its Fire Drill when every message's digest matches (details). Needs CLI
v0.0.51. - You can now restore a MongoDB database with a time-series collection, and drill it into a sandbox. Needs CLI
v0.0.51. - The daemon's warning when a host's clock goes backwards now gives the last backup's recorded time, where it could say 0001-01-01. Needs CLI
v0.0.51. - A rebuilt host that took over an old one with
safegrd enroll --token <node token> --node-id <id>backed up to a new key SafeGrd had no record of. It now backs up to the host's own key: a SafeGrd-managed key is fetched, and for a customer-managed key enroll asks for the key file (details). Needs CLIv0.0.51. - In a project that keeps weekly and monthly copies, a daemon writing to your own bucket now locks those copies, including ones from the last 45 days still in the bucket.
safegrd mcpnow answers requests piped into it before it exits, where it printed nothing when stdin closed straight after them. Needs CLIv0.0.51.- The Projects table no longer counts a host as one of the project's surfaces.
- A surface's details, under Storage, and the storage table under Projects no longer overcount incremental backups ("Backups in the last day: 8" for one).
- Billing now shows the current surface count and hosted storage, where a surface added since you opened the console was missing until a reload.
- The storage table under Projects no longer lists a host as a surface "not counted until its first backup".
- A console link to
#tokens,#surfacesor#drillsnow opens that section. safegrd doctor's fix for a role that can write now links the read-only role, where it linked a docs heading that does not exist. Needs CLIv0.0.51.- Cloudflare R2 has no S3 Object Lock, so set
worm_mode: "NONE"for it, as the docs now say (details). - SafeGrd now retries drill evidence that Vanta or Drata did not accept, for several hours.
- A MySQL backup through mysqldump 8.0 over TLS no longer prints
unknown variable 'loose-ssl=1'. - A backup in a directory on the host no longer records a lock, matching the console.
- A restore or sandbox drill of a PostgreSQL database backed up with
change_log: truenow leaves out the change log, whosesafegrdschema and triggers stay in the source. safegrd restore --tableof a backup taken withchange_log: trueno longer fails withschema "safegrd" does not exist.- A MySQL or MongoDB backup on a host without
mysqldumpormongodumpnow fails before it writes anything to storage. - After you rename a directory, the backup no longer counts its files as changed: they are read again with no new content.
- When a drill on SafeGrd finds part of a hosted backup damaged or missing, the surface's next backup now uploads everything again, as after a drill on your host.
- A backup that became the weekly or monthly copy after it was written unlocked now shows as locked until the copy's date, in the console, the Evidence Pack and
safegrd list. - In your own bucket, a project that locks only its kept copies no longer leaves a week without a copy by picking a backup
safegrd prunealready deleted.
17 October 2026: point-in-time recovery, backups at a time of day, and rows out of a backup
Needs CLI v0.0.49.
New
- You can restore a self-hosted PostgreSQL database to any second in its recovery window, not only to its last backup:
safegrd wal restore --to "2026-10-17 14:52:00", or--to latestfor everything archived. Setarchive_commandtosafegrd wal pushand the daemon does the rest. How the WAL is stored and locked is in the details. - Point-in-time recovery works in your own bucket on every plan, and on hosted storage on Growth and Scale for the window that the pricing page lists. On hosted storage SafeGrd sets the locks, so the host holds no storage key.
- You can mark a restore point before a migration with
safegrd wal point, and the daemon marks one each day with every table's row count. Every 7 days the daemon test-restores the archive into a throwaway cluster on the host. It restores to the newest restore point, checks each table's count, then restores to the end of the archive. - You can restore a SQLite database to within a second of any moment in the window too:
safegrd wal restore --to "..." --out restored.dbwrites the database as it was (details). The resident daemon follows the database's WAL every second and checkpoints it itself. Each day it records a restore point with every table's exact row count, and the drill checks a restore against it. - You can see a surface's recovery window in the console, and you get an alert when
archive_commandfails, a WAL file is missing or a point-in-time drill fails. The surface's certificate states the window and the last drill. - You can schedule a surface at a time of day (
at 02:00) or with a cron line (0 3 * * 1-5).scheduleis read in UTC unless the surface setsclock: local(details). The console shows the next runs as you type. - You can keep scheduled backups out of busy hours:
blackoutlists times of day when none starts, such as["22:00-02:00"]. A backup that falls due inside a window waits for it to end. Back up now is not held. - You can skip a backup when nothing changed:
skip_if_unchanged: trueon an incremental surface writes none. The console then shows the surface as unchanged since its last backup instead of Overdue. In your own bucket and on local storage. - You can get a few rows back without restoring their table.
safegrd verify --queryrestores the backup into a sandbox, runs one query in a read-only transaction and prints the rows as CSV (details). safegrd restore --tableworks on a full backup too.- You can store backups in a cheaper storage class:
storage.storage_classtakesSTANDARD,STANDARD_IAorGLACIER_IR. The cost calculator shows what each drill pays to read a backup back. - You can retry a failed
pre_backuphook:pre_backup_on_error: retryruns it again after 10 and then 20 seconds.post_drillruns a command after each Fire Drill the daemon runs, withSAFEGRD_DRILL_STATUSandSAFEGRD_SNAPSHOT_IDset (details).
Fixed
- A backup run by hand no longer prints that it is stored before it has run, and a failed upload says the bucket refused it.
- A sandbox drill finds each table's newest row again.
- The daemon logs when the host's clock steps.
- A SQLite checkpoint by the daemon that cannot take the write lock now waits for it instead of starting over.
- The console and the Evidence Pack no longer call a backup in a directory on the host locked: it shows as kept until its date on the host's disk.
- A surface SafeGrd backs up says when its next backup or drill waits for the organization's minutes on SafeGrd, and when it can run, instead of "due now".
- A surface under a host with no daemon says its backups run from your own scheduler, not "on the host's next check-in".
- A drill on SafeGrd that did not run says why.
13 October 2026: incremental MySQL and MongoDB backups
CLI v0.0.48.
New
- Your MySQL, MariaDB and MongoDB backups are now incremental, as PostgreSQL and SQLite backups are. After the month's first run, each run uploads only what changed in each table or collection, and every run is a restore point. On a host, in the daemon and for a database SafeGrd backs up. On a 270 MB MySQL database and a 290 MB MongoDB one, a run with nothing written uploads 0.1 MB, and 2,000 new rows upload 0.5 MB.
--format tar, or Full as the Backup type when you add the surface, keeps one archive per backup. - You can see what each MySQL table and MongoDB collection uploaded in the snapshot detail, as for PostgreSQL tables.
- You can script a backup:
safegrd backup --surface ID --jsonprints the snapshot's record on stdout, with the run's progress on stderr. - You can find a file by its name alone, at any depth:
safegrd find contract.txt.restore --pathtakes a bare name too. - You can check a Laravel app's setup with
php artisan safegrd:doctor: the connection, the CLI, its home and the schedule.safegrd:listlists the backups, andsafegrd:upgradereplaces the app's copy of the CLI with the latest release. A failed backup is written as an error to the app's log, where your error tracker sees it, and dispatched as aBackupFailedevent (details).
Changed
- You choose a Backup type, Incremental or Full, when you add a MySQL or MongoDB surface in the console. A surface added before the choice existed writes incremental backups from its next one.
- In a project that locks only the copies the plan keeps, MySQL and MongoDB backups now follow that rule, as PostgreSQL and files do. Full backups and mailboxes still lock every copy.
- The CLI says Locked until for every locked backup.
safegrd listgives both dates for a backup kept past its lock, as on the free plan. - In a kept-copies project, the Evidence Pack lists a recent backup kept without a lock as
unlocked, with the date it is deleted. It leaves that backup out of the locked copies it counts. safegrd export --to-bucketchecks that the bucket has Object Lock before it copies anything, and says how much it copied.- When the database refuses to sign in a backup SafeGrd runs, the error says to replace the connection URL. A backup you ask for while your plan's machines are all busy says it is waiting.
- A host enrolled without a surface, such as a machine set up only to restore, is listed under Hosts, not as a surface that never backed up.
- The Laravel package installs on Laravel 13, and its scheduled backup also runs while the app is in maintenance mode. It passes the connection's
MYSQL_ATTR_SSL_CA, and PostgreSQL'ssslrootcert,sslcertandsslkey, to the CLI. An app you rename keeps backing up as the surface it was set up as.
12 October 2026: Supabase Storage back into its bucket, buckets SafeGrd backs up, Laravel apps, and each table across backups
CLI v0.0.47, and v0.0.48 for the columns.
New
- You can follow each table across a database's last 30 backups in the Tables tab of its panel. It shows the row count, with the backups a drill restored it in, size, newest row, when its definition changed and what the last drill found. A row opens to the table's columns and what a restore of it alone needs (details).
- You can see Threat Shield's record on every completed snapshot, not only a flagged one: the backup it was compared with, each check with its numbers, and what changed between the two. That covers tables added and gone, definitions that changed, columns lost or gained, and the tables whose rows moved most (details).
- You get a
column_dropwarning from Threat Shield when a table has lost a column since the backup before. The rows stay and the column's data does not, so no row count shows it. Each PostgreSQL, MySQL, MariaDB and SQLite backup now records every table's columns for this. Needs CLIv0.0.48. - A failed sandbox drill now names the table that did not come back, as an in-memory drill already did. Needs CLI
v0.0.48. - You can restore a files snapshot straight into a bucket:
safegrd restore --target s3://bucket/prefix, with the same--source-*flags as the backup. Each file keeps the content type the backup recorded, and its hash is checked against the backup as it uploads. After a Supabase incident, restore the database, then the Storage files into their bucket, with no second tool (details). Needs CLIv0.0.47. - You can back up a Laravel app as one surface:
type: laravelwith the app's root. Each backup reads the connection from.envand takes the database (MySQL, MariaDB, PostgreSQL or SQLite), the.envandstorage/appas one snapshot, uploading only the files that changed. The drill also checks that.envholds a usableAPP_KEYand that themigrationstable agrees with the migration files.safegrd restore --target … --target-dir …puts the database and the files back (details). Needs CLIv0.0.47. - You can set a Laravel app up two ways. Run
composer require safegrd/laravel, thenphp artisan safegrd:install, which installs the CLI, signs in and lets Laravel's scheduler back the app up and drill it. Or runsafegrd init --laravel /var/www/appon a host with the CLI. The console's Add surface shows both.safegrd daemon run --once --interval 5mfrom a scheduler now counts as a daemon, so it takes Back up now and Drill now from the console. Needs CLIv0.0.47. - You can have SafeGrd back up a bucket for you, with nothing installed: Supabase Storage or any S3-compatible store. Under Back up on SafeGrd, give its endpoint and an S3 key. SafeGrd holds this bucket's key, encrypted, and uses it only on the machine it starts for each backup. The backups go to hosted storage and are drilled on your plan's cadence (details).
- You can see each MongoDB collection's top-level fields and their BSON types in the Tables tab. Threat Shield warns with
column_dropwhen a field is gone from the next backup (details). Needs CLIv0.0.48. - You can have SafeGrd skip tables nothing wrote to since the last backup of a PostgreSQL database it backs up, with a change log. Tick its box in Back up on SafeGrd or the surface's Edit dialog. It installs a
safegrdschema and one trigger per table in your database, so it stays off until you do. Not offered with the Supabase sign-in, whose role is read-only. - You can switch a PostgreSQL surface SafeGrd backs up from a typed connection string to a platform sign-in. Sign in through a platform on its Details asks which of Supabase, DigitalOcean and Heroku. A surface that signs in already offers Reconnect there, as before.
- A backup through the role a Supabase, DigitalOcean or Heroku sign-in gave it now stops when row-level security hides rows from that role. Before, it recorded the rows it could see as a complete backup. The error names the tables and what to do. A role you typed yourself is backed up with a warning, as before. Needs CLI
v0.0.48.
11 October 2026: sign-in methods in Settings, and a CLI sign-in that asks for a code
CLI v0.0.45, and v0.0.46 for the new sign-in.
New
- You can restore rows into a hosted Supabase project with
--data-only-schema. The restore leaves out Supabase's own migration tables, which thepostgresrole may not write, names each with its row count and loads the rest. Before, the first of them ended the restore with no user put back. Tested from one hosted project into a second, empty one (details). Needs CLIv0.0.46. - You can back up a bucket over the S3 API, as you would a directory. Run
safegrd backup --files s3://avatarswith the store's endpoint and keys, or add a surface with anobject_storeblock. It is for Supabase Storage, whose own backups leave the uploaded files out. Each object keeps its key and content type, and restores into a directory (details). Needs CLIv0.0.46. - You can see and unlink the Google and GitHub accounts linked to yours on the Sign-in methods card in Settings. Sign out everywhere ends every console session, this one included, and makes every remembered browser ask for its code again. Access tokens keep working (details). You cannot unlink the link that made a password-less account, or a provider your organization requires.
- The first time you sign in with Google or GitHub, it links to an existing account with your address only where the provider owns that address. That is Google, for a Gmail address or a Workspace account of the address's domain. Any other address gets a mailed link, and the account links when you press that page's button.
Changed
safegrd loginprints a link and a short code. The page asks for the code, and shows the address and client that started the sign-in, so a link forwarded to you approves nothing on its own. The WordPress plugin signs in the same way. Needs CLIv0.0.46and plugin 0.1.7.- Creating an access token or authorizing a CLI sign-in now takes a single-use 6-digit code mailed to you, or your authenticator's code. A fresh Google or GitHub sign-in no longer stands in. With two-factor on, creating a token or authorizing the CLI asks for the authenticator's code, as deleting one does.
Fixed
- With two-factor on, a second Google or GitHub sign-in now asks for your code before it moves between the sign-in page and the CLI authorize page.
- A Google or GitHub sign-in no longer claims a password account whose address was never verified. The sign-in is refused and the page says how to claim the account by its inbox.
- A restore from hosted storage no longer fails with Rate limit exceeded: it waits and tries again, as a backup does. Needs CLI
v0.0.45. list,verifyandrestorenow find the new snapshots of a host that backed up before it enrolled. Needs CLIv0.0.45.
10 October 2026: each surface's last 30 days, Hosts, and one name per thing in the CLI
CLI v0.0.43. Its config file and some commands changed: rewrite a config before upgrading a host to it.
New
- You can add a DigitalOcean or Heroku PostgreSQL database with nothing typed: Back up on SafeGrd signs in to the platform and lists your databases. SafeGrd reads the connection from the platform for each backup, so a password reset or a Heroku credential rotation needs nothing from you. On a DigitalOcean cluster that lists trusted sources, SafeGrd adds the addresses its machines connect from (DigitalOcean, Heroku).
- You can choose which of a new project's backups are locked: every backup, or only the kept copies. With kept copies, the day's first backup is locked when written, and the weekly and monthly copies are chosen from backups a drill restored. The backups in between are kept unlocked and count at half against hosted storage. The organization's first project locks every backup, and a project keeps its rule (details). Needs CLI
v0.0.44for a kept-copies project. - You can see each surface's last 30 days, one mark a day: restored by a drill, backed up, no backup, failed, or nothing due. Each day is judged by the schedule the surface had that day.
- Overview shows every surface's 30 days in one grid. Hosts lists each enrolled host with the surfaces it runs, its daemon, last check-in and CLI version (details).
- Details opens a panel beside the list with the surface's actions, where its data goes, the last 48 hours with each drill drawn back to the snapshot it restored, and its snapshots and drills.
- Fire Drills has a calendar of each surface's drills, and each view has its own address you can bookmark.
- You can restore single plugins, themes or tables from a WordPress backup and leave everything else as it is. Each is listed with the version in the backup and the one installed now.
wp safegrd contentslists them, andwp safegrd downloadwrites single tables, plugins or themes to a file. The Updates, Plugins and Themes screens show how old the last backup is, with Back up now (details). Needs plugin 0.1.6. - You can undo a WordPress restore with Put the copy back, or
wp safegrd restore --undo. The plugin's Restore & download tab opens one backup at a time, with each part's size and versions, and shows what will be replaced before anything starts (details). Needs plugin 0.1.6.
Changed
- The free plan drills each surface once a week, in memory, instead of once a month.
- Surfaces lists surfaces only, each with one word for its state. Chips count surfaces by state and filter the list, and the list groups by project, by what runs the backup, or by host.
- Snapshots are grouped by surface, and the ones a drill restored are marked.
- The config file has
server:andnode:blocks, and every backup source is a surface. A key that moved is refused with the key that replaces it (the file). Needs CLIv0.0.43. - One name per command:
safegrd verify historyreplaceshistoryandverify-history,project listreplacesprojects,verify --in-memoryreplaces--dry-run, andenrolltakes--tokenonly.safegrd doctorreplacesstatus, fails a refused token, names a newer CLI and reads the storage (commands). Needs CLIv0.0.43. - The GitHub Action at
safegrd/backup-action@v0calls the new names. With neitherdatabase-urlnorsurface, it backs up the config's one surface.
Fixed
- A backup run with one-off flags on a config with one surface now backs up that surface, where it started a second history for it. Needs CLI
v0.0.43. - Retention now follows a surface's own storage block. Needs CLI
v0.0.43. safegrd config validatesays why it refused a file. Needs CLIv0.0.43.- Drill now is no longer refused as not complete in hosted storage when a surface's two newest backups started in the same second.
- A WordPress backup or restore that the host kept stopping at the same point no longer starts again without end: it fails and says what it was doing. A restore the host stops partway now carries on, and a restore of every plugin keeps the SafeGrd plugin in place. Needs plugin 0.1.6.
7 to 9 October 2026: storage forecast, Help in the console, and Supabase tables a role cannot read
CLI v0.0.33 to v0.0.40.
New
- You can have Fire Drills run your own checks after the restore: a query and the value it must return, or a command that must exit 0. Set queries under Drill checks on a database surface in the console, queries and commands under
drill.checksin the host's config, or checks for one drill withsafegrd verify --checks. Drills on SafeGrd run the console's queries. Each result is on the drill record and covered by its certificate hash (details). Needs CLIv0.0.42. - You can see on the Storage tab what a project's storage holds now and what it will hold once every surface has run through its retention. On hosted storage it shows the day it reaches what your plan includes (details).
- The organization's half of the Storage tab is headed Organization storage, and the key access log has a Project column.
- You can find the guides for what your organization backs up under Help in the console, with how to reach support. Copy diagnostics there copies what support needs and no token, key or password (details). Each surface's Details links its guide.
- You can make an age key pair for a customer-managed key with
safegrd keygen, with no other tools installed. Needs CLIv0.0.40. - You can drill one host's snapshot from another host in the same organization, and the drill is recorded on that surface. A host enrolled only to restore can drill the backups of a host with a customer-managed key (details).
- You can see in wp-admin when SafeGrd next test-restores your WordPress site, and how often your plan does. The WordPress docs cover restoring some parts, downloading a backup, leaving files out, the backup hour and logs, with two short clips.
- Backup pages for PostgreSQL, Supabase, Neon, Railway, Render, Fly, DigitalOcean, Hetzner, Amazon RDS and Heroku, each with the manual steps; a list of PostgreSQL backup tools, each with what it does best; integrations; and two guides, pg_dump, pg_basebackup or pgBackRest and RPO and RTO.
Changed
- Surfaces has one Add surface button. It asks first where the backup runs: on your server, or on SafeGrd for a managed database such as Supabase, Neon or RDS. On your server it has three steps: what to protect, how to set it up, and the host that runs it.
- Back up now reads Backup queued, Backup starting or Backing up and is off while a backup runs. A queued or running backup or drill shows on the surface, and a failure shows once, under Status.
- A snapshot's Details lead with how to restore it, then the commands.
- The two-factor card offers no Turn off while your organization requires it, and asks no password of an account made by a Google or GitHub sign-in.
safegrd prunewrites deletions to stdout and warnings and errors to stderr. Needs CLIv0.0.38.- The CLI is built with Go 1.27.2, which fixes twelve standard library vulnerabilities in 1.27.1. Needs CLI
v0.0.39.
Fixed
- A PostgreSQL backup no longer fails on a table its role cannot read, which stopped every Supabase backup through the sign-in role. The backup leaves the table out, with its partitions, sequences, and the views and row-level security policies that read it, and names each in a warning. A restore leaves out the foreign keys into it, and the snapshot's drill passes (details). Needs CLI
v0.0.39. - A Supabase backup through the sign-in role reads as
postgreswhere the project allows it, as Supabase's own CLI does. Needs CLIv0.0.37. - A MySQL account with
SELECT,SHOW VIEW,TRIGGERandLOCK TABLESno longer fails onSHOW EVENTS, and backs up without events (the grants). Needs CLIv0.0.37. - A host with 2.6 GiB free of 72 no longer refuses to drill a 137 KiB snapshot: the CLI keeps 5% of the filesystem free, at most 1 GiB. Needs CLI
v0.0.37. - A backup to hosted storage no longer fails with Rate limit exceeded: it waits and tries again. Needs CLI
v0.0.40. - A WordPress site now shows as WordPress and says the plugin runs its backups, where it showed a Postgres badge and said no daemon ran it.
7 October 2026: joining an organization takes an accepted invitation
- When you invite someone, they get an email and join when they accept it. Before, an address that already had an account became a member at once. The link works for 7 days, and a new address creates its account on the link's page (details).
- You can see invitations not yet accepted under Team members, with Resend and Revoke.
7 October 2026: Supabase backups with no password typed, and what runs next
- You can back up a Supabase project on SafeGrd by signing in through Supabase instead of typing a password. SafeGrd holds the sign-in, encrypted, and each backup's machine gets only a read-only role whose password expires within minutes. You can still type the password, and switch between the two from the surface's Details. A sign-in is refused when row-level security hides rows from its role, with the tables named (how it works).
- Every surface's Details say when its next backup and drill are due, and who starts each: SafeGrd, the host's daemon at its next check-in, or your cron or workflow.
- The Fire Drills tab has Upcoming: each surface's next drill, overdue ones first, where it runs and how often the plan drills (details).
6 October 2026: WordPress sites, backed up and restored from a plugin
The SafeGrd Backup plugin backs up a WordPress site from wp-admin, with no SSH and nothing installed beside WordPress, so it works on shared and managed hosting. It is free for one site. WordPress has the details.
New
- You connect a WordPress site from Tools, SafeGrd: sign in or create an account in the tab that opens, and the first backup starts. With WP-CLI, run
wp safegrd connect --token=env:SAFEGRD_TOKEN. - Each backup is a complete snapshot of the database and
wp-content, encrypted on the site's own server, and uploads only what changed since the last one: an unchanged image, theme or table uploads nothing. - You can back up a site on a host that stops long requests: the backup picks up where the host stopped it.
- SafeGrd test-restores the site in memory on its own machines: every table's rows, every file's SHA-256, and every attachment the database names.
- You can restore from wp-admin onto a new install or over the site itself. Nothing changes until every table and file is loaded and checked. Then they are swapped in, and what they replace is kept until you delete it. A new address is replaced in the database, serialized values included, and a different table prefix is handled.
safegrd restore --target mysql://… --target-dir …restores a WordPress snapshot from the command line.
5 and 6 October 2026: MySQL, MariaDB and MongoDB with no server, and fixes from a run on production
We ran eight customer journeys against safegrd.dev with the released CLI, from sign-up to restore, and
fixed what they found. CLI v0.0.22 to v0.0.31.
New
- You can have SafeGrd back up and drill a MySQL, MariaDB or MongoDB database with no server of your own, each drill in a sandbox of its own engine. The MySQL and MongoDB pages say which connection string to paste, and the TLS options say what
tls=true,tls=skip-verifyandssl-caeach do. safegrd restore --tablebrings back a table the database no longer has, created from the backup's own schema.- Roles, owners and grants restore into a fresh PostgreSQL cluster.
- Every snapshot has a recovery document beside it, so you can restore it without SafeGrd.
safegrd keepextends one snapshot's lock,safegrd estimateprices a schedule's database egress before the first backup, andsafegrd findsearches every surface this host holds a key for.- The billing page shows the card renewals charge, lets the owner update it, and lists invoices with a PDF.
- An invite emails the person a single-use link naming the organization and the role.
- The CLI's exit code says which side failed: 10 the database or source, 11 SafeGrd, 12 storage.
Changed
- A machine enrolled only to restore no longer counts as a protected surface, so on the free plan you can enroll a second machine to test a restore or replace a lost host. A host counts once it stores a backup of its own.
- A restore asks SafeGrd for the one key its snapshot needs, and each release is recorded.
enrollasks which project when the organization has more than one, and tells a host with nothing to protect how to restore.- Alert email is off until an owner or admin turns it on, under Settings.
- A failed drill shows on the surface with its reason.
Fixed
- After a failed backup on SafeGrd, an hourly surface now tries again after one interval, or sooner when SafeGrd caused the failure, instead of six hours later.
- Back up now is no longer refused for an hour after a failed backup, so you can check a fixed connection string at once.
- A failed MongoDB backup now shows the server's answer, such as authentication failed, where it gave no reason.
- A database tool's error no longer repeats the connection string's password: it is replaced before an error is printed or stored.
- The console now says a run that stored nothing failed, where it said the snapshot's lock had ended.
- The remote MCP server no longer refuses every request.
- A project-scoped token no longer lists every project's surfaces.
4 October 2026: incremental PostgreSQL and SQLite backups
New
- You can check every surface before its first backup:
safegrd doctoropens each with the credential its backup will use, including one SafeGrd holds for you. On an enrolled host the console then shows the host as Checked, or what failed. The setup page's host step asks you to run it before the first backup. Needs CLIv0.0.14. - SafeGrd tries a connection string when you enter it for a database it backs up. A wrong password, an unknown database or a host that does not answer shows as a warning in the database server's own words. The surface is still added, since a backup connects from a different machine than this check.
- Your PostgreSQL backups are now incremental: after the month's first run, each run uploads only what changed in each table, and every run is a restore point.
--format tarkeeps one archive per backup. Needs CLIv0.0.12. - You can restore chosen tables from one backup into a running database with
safegrd restore --table, and list a table's versions withsafegrd find --table. - You can skip reading tables nothing wrote to since the last run:
safegrd backup --change-logtracks writes with a trigger on each table (how it works). - The console's table breakdown shows what each table uploaded in a run.
- Backups SafeGrd takes of a PostgreSQL surface are incremental too.
- Your SQLite backups are incremental as well: a run uploads only the parts of the database that changed. Needs CLI
v0.0.12. - You choose a Backup type, Incremental or Full, when you add a PostgreSQL, SQLite or files surface in the console. Full is one archive per backup.
3 October 2026: drills on every plan, incremental file backups, two-factor sign-in
Plans
- Every plan runs Fire Drills. Free drills each surface once a month, in memory. Starter drills weekly and adds sandbox drills and the remote MCP server. Growth and Scale drill daily. Prices are unchanged.
- On paid plans, hosted storage past the included amount is billed per GB-month instead of refusing backups. The rate is on the pricing page.
New
- Your file backups are incremental: after the first run of the month, a run uploads only the files that changed. You can bring back one version of one file with
safegrd findandrestore --version. Needs CLIv0.0.10. - You can have SafeGrd back up a database with no server of yours beside it, such as Supabase, Neon or RDS. Give the connection string under Back up on SafeGrd on Surfaces, and SafeGrd keeps it sealed. A machine SafeGrd starts for each backup dumps the database, encrypts it and writes it to hosted storage, and drills follow on your plan's cadence. PostgreSQL, MySQL and MongoDB, on paid plans and the trial.
- You can have SafeGrd run a surface's drills, so your host needs no database server to restore into: Drill on SafeGrd in the console. PostgreSQL, MySQL, MariaDB, MongoDB and SQLite restore into a throwaway database, and files and mailboxes are read back in memory. Each drill runs on its own machine, deleted when the drill ends. For surfaces with a SafeGrd-managed key and backups on hosted storage.
- Sign in with Google or GitHub.
- You can turn on two-factor sign-in with an authenticator app. A browser can be remembered for 30 days, and an organization can require the code at every sign-in.
- You can take a locked snapshot before a destructive command with
safegrd guard. /docs/agents has hook recipes for Claude Code, Cursor and Codex, andsafegrd doctor --agent-proofchecks what an agent on the host could delete. - You choose key custody per host. A new host gets a SafeGrd-managed key unless you enroll it with
--key-custody local(orSAFEGRD_KEY_CUSTODY=local) for a customer-managed key, and the installer no longer asks. Each surface in the console says which mode its key is in, and the organization page no longer has a custody setting. - A PostgreSQL sandbox drill needs no setup on a host with PostgreSQL installed: the daemon starts a temporary server and removes it afterwards.
- You can run the CLI as a signed container image,
ghcr.io/safegrd/cli, with a Compose file, a Kubernetes CronJob and a Helm chart. - You can back up and drill a database, a folder or a mailbox from GitHub Actions with
safegrd/backup-action, with no server of your own. - You can download a backup on hosted storage from the console, under Telemetry. It arrives encrypted, with its metadata file, and
safegrd restore --fromrestores it. - You can delete an organization from Settings.
- Trust and security, compliance, about, security.txt and this changelog. Sample drill records and a sample Evidence Pack are on the attestation record page.
Changed
- Setup asks how your key is managed. With a SafeGrd-managed key, the default, SafeGrd keeps your key sealed and releases it only to your enrolled hosts. The terms and sub-processor list describe both modes.
- You confirm your email address with a code. An account that has not yet is sent one at its next sign-in.
- The console says why a drill ran in memory or did not run, for example a host short of disk space.
- The Evidence Pack states the evidence for each control and says it is not a certification.
- The sub-processor list, the DPA and the privacy notice add Fly.io, which runs drills on SafeGrd.
Fixed
- A member with the viewer role can no longer enroll a host.
- A replacement host can now restore a lost host's snapshots under a SafeGrd-managed key.
- A database that does not answer now fails the backup after 30 seconds instead of holding it open.
safegrd daemon installnow starts the service, anduninstallstops it.- A README badge for a drill under an hour old no longer shows as a broken image.
29 September 2026: the daemon, agents and Supabase
- You can run the daemon on a host: it runs every surface on its schedule and checks in with the console.
safegrd claimadds surfaces named in the console to a host that is already enrolled.- Gmail is backed up once with its labels, and a mailbox restores from an export.
- A guide for backing up and drilling Supabase from GitHub Actions, with no server of your own.
- CLI
v0.0.6tov0.0.9.
28 September 2026: guided setup and custody
- Setup is two short wizards in the console: storage, then the surfaces to protect.
- You can have SafeGrd keep a surface’s credentials and your Age key sealed, released only to your enrolled hosts.
- The storage step writes a least-privilege bucket policy and AWS templates that create the bucket locked.
- Every backup and drill request records who asked, and how.
- CLI
v0.0.5.
25 September 2026: hosted storage, MCP and Fire Drills from the console
- Hosted storage: you can back up into SafeGrd’s locked bucket, with no bucket of your own. Your hosts never hold a storage key.
- The remote MCP server: an AI agent can read backups and drills and ask for a backup or a drill. No agent token can delete one.
- Ask for a Fire Drill from the console, and read its record there.
- CLI
v0.0.1tov0.0.4.