Docs
Short guides to everything in Dblore.
Install
- Download the signed and notarized DMG from Releases, drag Dblore to Applications.
- Needs macOS 14+ on Apple Silicon.
- Updates itself: a green update button shows when a new version is out; Dblore > Check for Updates… checks now.
- Opens
.sqlite,.sqlite3and.dbfiles from Finder, and.duckdbfiles once the DuckDB plugin is installed.
Connect
- PostgreSQL: host, port, user, database. SSL modes from
disabletoverify-full, timeouts, retry on failure, read-only mode. - MySQL and MariaDB: host, port, user and database (required; the sidebar shows only that database). Direct or through an SSH tunnel. Users with
caching_sha2_passwordormysql_native_password;sha256_passwordand MariaDB ed25519 are not supported. - MySQL SSL: only Verify CA and Verify Full check the server certificate. Allow, Prefer and Require encrypt the connection without checking it, so on an untrusted network an active attacker could read the traffic and the password. Allow and Prefer also fall back to plain text when the server offers no SSL. A client certificate needs Require, Verify CA or Verify Full.
- MySQL full password: when the server asks for the full password (
caching_sha2_passwordwith no cached login) and the connection does not verify the server's certificate, Dblore stops instead of sending it and says to choose Verify CA or Verify Full. A direct connection to127.0.0.1,::1orlocalhostis exempt. - SSH tunnel (PostgreSQL, MySQL, MariaDB, MongoDB, SQL Server, ClickHouse and Redis; MongoDB takes a single host, no
mongodb+srv://): password or ed25519/ECDSA private key (RSA is not supported). A key passphrase is never stored. The first connection asks you to trust the host key. A changed host key blocks the connection and shows the expected and the presented fingerprints: once the server's administrator confirms the new one, Replace Pin (never the default button) stores it and connects again. MCP, the command line and Shortcuts never replace a pin. - Client certificate: authenticate with a certificate, keep its passphrase in the Keychain and pick a custom certificate authority. A custom CA also works on its own, without a certificate.
- Statement, lock and idle-in-transaction timeouts: global in Settings > Results, overridable per connection. PostgreSQL, MySQL/MariaDB, direct MongoDB (
directConnection=true) and SSH tunnel sockets use TCP keepalive (30 s idle, then 3 probes 10 s apart), so a dead link shows as a lost connection instead of hanging. - Passwords are stored in the Keychain. When the Keychain cannot read a saved password (locked or access denied), Dblore stops instead of connecting without it. Ask on connect (Password source in the connection form) asks for the password at every connect instead: it is kept in memory for that session's reconnects and never saved, saving the connection removes a stored one, and MCP, the command line and Shortcuts refuse the connection.
- MongoDB: a
mongodb://ormongodb+srv://string, or the form. See MongoDB. - SQLite: pick a file in the connect modal. Features the engine cannot support are hidden.
- DuckDB: needs the plugin.
- SQL Server: needs the plugin. Default port 1433. Verify Full checks the certificate and the host name; Read-only is checked by Dblore, so for a guarantee on the server use a login with only
db_datareader. - ClickHouse: needs the plugin. HTTPS on port 8443 by default; turning SSL off moves a port still at 8443 to 8123 (HTTP). A Read-only connection is enforced by the server.
- Redis: needs the plugin and Redis 6.0 or later. Host, port (6379), an optional ACL user, the password and a database index from 0 to 15. TLS is Verify Full (checks the certificate and the host name, optionally against a custom CA) or Disable; plain TCP goes only to a host on the local network (
localhost, a.localname or a private address) or through an SSH tunnel. Commit styles are Immediate, Confirm and Password (Redis has no transactions, so no Review). - Presets: the engine picker lists CockroachDB (port 26257), Redshift (5439) and YugabyteDB (5433) under PostgreSQL, and TiDB (4000) and SingleStore (3306) under MySQL. They connect with that driver and its default port; the PostgreSQL presets also leave out the session settings and catalog reads the engine lacks. Redshift and SingleStore are untested. An older Dblore opens a preset connection as plain PostgreSQL or MySQL.
- Import Connections… (welcome screen or File menu) reads
.pgpass, a PostgreSQL,mysql://ormariadb://URI, DBeaver (with saved passwords), TablePlus (Redis connections included) and DataGrip, and a Dblore connections file. Preview and pick before importing; duplicates are flagged. - Export connections: the welcome screen's Connections tab exports chosen connections, and optionally an open workspace's favorites, to a Dblore file. Passwords, SSH secrets, keys, certificates and file databases are never exported. On import, connections already saved are skipped, a connection string pointing at other hosts is dropped, SSH tunnels and weaker settings are flagged, and favorites go into one folder without duplicates.
- Scripts: a connection can run your own scripts and get its password from a command. See Connection scripts.
- Vault: a connection can get its login from HashiCorp Vault on every connect. See Vault credentials.
- The welcome screen lists saved and recent connections (up to 50), each with its endpoint (
host:port@user, plusviathe SSH host, or the file name for SQLite and DuckDB); search matches it. Hover a card to remove it. - Groups and favorites: the Group field of the connection form takes a path such as
Prod / EU / Web(up to three levels; deeper ones merge into the third), shown as nested, collapsible groups on the welcome screen and in the form's history menu. The star on a card adds it to Favorites, listed first and never trimmed from the 50-entry history; connection exports keep groups and stars. - Masking (SQL connections): rules in the connection form hide matching columns as Full, Last 4, Email or NULL in results, copies, exports, AI, MCP, Shortcuts and the command line. A rule is
column,table.columnorschema.table.column(*only for a schema or table). Masked cells cannot be edited and show a mask glyph in the header, the export sheet keeps them checked, and the Data Rewind preview shows mask text while Rewind restores the real values. A rule change applies from the next query, also in other windows. Masking is not access control: it matches column names, so an alias or a computed column gets past it, filters and Find in every table still search real values, and dumps, backups and Copy Objects keep them. While a transaction is open, a table renamed by another session since this connection last read it is matched under its old name until the transaction ends. - The footer shows the connection name, the Safe Mode cues and a No SSL badge when the connection is not encrypted. A MySQL connection encrypted without a checked certificate shows Unverified SSL.
- Attached databases (SQLite): the connection form attaches up to 10 other SQLite files, each under its own name, on every connect. Read-only connections attach them read-only, a missing file is skipped with a warning, and a locked file asks to unlock first. Their schemas show in the sidebar and autocomplete. Connection exports never include them.
Connection scripts
- The Scripts section of the connection form picks up to three scripts: Before connecting runs first, and a failure stops the connect; Password command prints the password on its first output line, used for that connect only and never saved (engines that use a password); After disconnecting runs once the connection is closed, or when the connect fails after the first script. Quitting Dblore runs the after-disconnect scripts.
- Scripts live in
~/Library/Application Scripts/ace.thi.Dblore, the only folder macOS lets Dblore run your scripts from. Open Scripts Folder opens it in Finder, creating it when missing. A script is a plain file name in that folder: anything that resolves outside it is refused. - Arguments: the connection's host, port, user and database, in that order. Quote them in your script and pass
--before them to other commands. A host, user or database that starts with-or holds a control character is refused, and the script does not run. - Script timeout: 1 to 300 seconds, 30 by default. A script that times out keeps running: macOS gives Dblore no way to stop it.
- Scripts run with your own environment, outside the app sandbox. Their output is never logged or stored, since a password command prints a secret. Only a before-connect failure shows the start of its error output, in the message to you; logs see the script name and exit status only.
- Imported connections never bring scripts with them: choose them in the connection form.
Vault credentials
- Password source in the connection form (engines that use a password): Saved password, Password command, Vault or Ask on connect (see Connect). Vault hides the password field and asks for the Vault address, the database secrets mount and role, and Token or AppRole (role id and secret id). Vault connections are always remembered.
- The address must be
https, or plainhttpto127.0.0.1,::1orlocalhost. Redirects are refused. An optional CA certificate (PEM) makes it the only trusted authority for Vault. - The token, role id and secret id stay in the Keychain, bound to that Vault address, mount, role and database target: change one and Dblore asks for the secret again instead of sending it elsewhere. They are never in the connection, an export, a workspace file or the logs, and are listed in Settings > Data.
- Test Vault takes a lease, shows its user and lifetime and revokes it at once. The token's policy needs
updateonsys/leases/revoke(besides Vault's default policy), or Dblore cannot end leases. - Every connect takes its own lease; the leased user and password serve that session only, and the saved user name stays. Dblore renews the lease at two thirds of its lifetime and retries a failed renewal until the lease really expires. When it ends, the connection closes and shows Vault lease ended with Reconnect, which takes a new lease; Dblore never reconnects on its own.
- Disconnect and quit close the session, then revoke the lease. Side connections (Find a Value, dumps, Copy Objects) take and revoke their own.
- The MCP server, the command-line tool and Shortcuts refuse Vault connections.
- A connections file keeps the address, mount, role, auth method and CA, never a secret. On import the address is shown in full and the CA by its SHA-256 fingerprint: pick the CA file and type the secret in the form. A workspace file never picks the Vault address; the saved connection's settings win.
- Older Dblore versions connect a Vault connection without a password, so it fails to authenticate.
Workspaces & tabs
- A workspace (
.sqlws) holds your tabs: notebooks,.sqlfiles, Markdown notes and table data viewers. - Several tabs and windows. Drag to reorder, middle-click to close.
- Drag a tab outside the window to open it in a new window (same as Open in New Window; pinned and unsaved tabs stay).
- Switch with Cmd+1 to Cmd+9. Reopen a closed tab with Cmd+Shift+T.
- Optional native macOS window tabs: enable Settings > General > Open new windows as tabs. Cmd+Shift+N or the + button opens a new window tab, Ctrl+Tab switches, Cmd+T opens a new tab. The header stays in the title bar, and the window tab bar sits below it, starting after the left sidebar; each window tab keeps its own Dblore tabs.
- Drag a window tab out of the tab bar (or use Window > Move Tab to New Window) to split it into its own window.
- Tabs and window state are restored on relaunch. Settings > General chooses between the welcome screen and the last session at startup.
- Command palette: Cmd+K finds and runs commands (not over Settings or the connection form).
Notebooks
- Notebooks are
.dblorefiles made of SQL cells, with inline results under each cell. - Run the cell with Ctrl+Enter, run and move on with Shift+Enter, run and insert below with Option+Enter, run all with Cmd+Shift+Enter.
- Cancel a running query. On PostgreSQL the server stops the statement and the connection is kept, with its temp tables and settings; an open transaction is rolled back. If the server cannot stop it, the connection is reset instead. Each result shows row count and execution time.
- New cell Cmd+Option+N, duplicate Cmd+D, delete Cmd+Delete. The floating toolbar of a cell has a Format SQL button.
- Named parameters: write
:namein a cell and fill the values in the inline form under it. You choose whether the values are saved in the file. - Old
.sqlnbfiles: rename them to.dblore, the content is unchanged.
SQL editor
- Open a
.sqlfile in a tab (Cmd+Shift+J for a new one). - For quick notes, Cmd+Shift+M opens a Markdown note in the workspace. It is written to a
.mdfile only when you press Cmd+S; the eye button in the header switches between the source and an editable preview. - Run the query with Cmd+R or Cmd+Enter. Query > Focus SQL Editor (Cmd+Option+1) and Focus Results (Cmd+Option+2) move the cursor between the editor and the result grid.
- Named parameters (
:name) are bound from the right sidebar. - Table- and column-aware autocomplete: Up / Down to choose, Enter or Tab to accept. It offers CTE names and their columns, after
ORDER BYorGROUP BYthe select list's aliases first, and after a join'sONthe foreign-key condition between the joined tables. Table and column names that contain what you typed follow the ones that start with it (ordfindspurchase_orders). - Bookmarks: Cmd+F2 or a click in the line number gutter toggles a bookmark on a line; F2 and Shift+F2 jump to the next and previous one, wrapping around. Bookmarks follow edits and are saved with the workspace. SQL editor tabs only, not notebook cells.
- Toggle comment with Cmd+/. Format SQL, word wrap and syntax highlight toggles sit in the editor toolbar.
- Settings > Editor > Formatter sets how Format SQL writes: indent (spaces or tab), keyword and function case, trailing or leading commas and line width, with a live preview. The defaults keep the usual output, and a format that would change the query's tokens leaves the text as it was.
- View > Zoom In (Cmd+= or Cmd++), Zoom Out (Cmd+-) and Actual Size (Cmd+0) resize the focused SQL editor, result grid or schema diagram. Editor and grid zoom change the global font sizes; the commands are off in notebook tabs.
- Explain with Cmd+E, Explain Analyze with Cmd+Shift+E (runs the statement, so it asks first).
- Snippets: type a trigger word and accept it from autocomplete. See SQL snippets.
SQL snippets
- Write snippets in Settings > Editor > Snippets: Add snippet, then a name, a trigger and a body, with a preview. Up to 200 snippets.
- Trigger: 1 to 32 letters, digits, underscores or hyphens, unique. Type it in a SQL editor; autocomplete offers the snippet, and accepting it inserts the body in place of the trigger.
- In the body,
${1:default}is tab stop 1 with its default text,${1}an empty tab stop,$0where the caret ends (the end of the text when there is none) and\$a literal dollar sign. A bare$1stays as written, so PostgreSQL parameters are safe. - Tab moves to the next tab stop and Shift+Tab to the previous one, in number order. Reaching the final caret, or moving the caret out of the current stop (a click, an arrow key), ends the snippet.
- Snippets are local data: Settings > Data exports and clears them with the rest.
Results grid
- Several statements give one tab per result, such as "1 · 120 rows" or "2 · UPDATE 3", in the editor and in notebook cells; from seven results the first five stay tabs and More holds the rest.
- Sort, resize (double-click a divider to auto-fit), search and paginate. The
#column stays pinned at the left edge when you scroll sideways. - Edit values inline. Changes are staged and committed as one batch, or applied directly depending on Safe Mode. Undo staged changes from the Edit menu.
- PostgreSQL views: cells of an auto-updatable view are editable when its key maps to its base table's primary key, staged and rewound like a table's (a row the edit moved out of the view cannot be rewound). Delete Row, Add Row, Duplicate and paste are refused on views; views with INSTEAD OF triggers or without such a key, and MySQL/MariaDB views, stay read-only.
- Foreign keys: look up the referenced row while editing, or jump to it.
- Pin result, re-run, then Compare: added, removed and changed rows are highlighted. Two
EXPLAINplans compare node by node. A notebook cell's pin is saved with its results. - Column header filter, plus a filter sidebar that builds
WHEREconditions. Filters can be saved. - Highlight matching cells or rows from the sidebar.
- Open a cell in the detail panel, with a JSON viewer and string beautifier.
- Show Row Details (the details button in a row's
#cell, or the right-click menu of a cell or of#) opens the whole row in the right sidebar, one field per column. In an editable result, change values there: edits stay drafts, marked with a dot, until Apply; Cancel drops them. Key, generated and read-only columns are not editable. - Add Row in the grid menu opens a form in the right sidebar: one field per column, a NULL button, checked against the column type. An empty field is left out of the
INSERT, and Add stages the row. - Settings > Results > Color Values by Type shows numbers, text, dates, booleans and JSON in their own soft colors.
- View Query shows the exact SQL sent, including the row limit.
- Numbers and dates: NUMERIC and DECIMAL values (PostgreSQL, MySQL/MariaDB, DuckDB) show, sort, sum and edit exactly, PostgreSQL NaN and Infinity included. A timestamp without a time zone (PostgreSQL
timestamp, MySQLDATETIME, DuckDBTIMESTAMP) shows its stored wall clock; a zoned one shows Mac local time with its offset; both keep fractions of a second. Copied rows and CSV, TSV, JSON and Excel exports write the wall clock, or ISO 8601 UTC for zoned columns. - Several cells at once: select rows, then right-click a column. Set Value… writes the same text, NULL or DEFAULT (PostgreSQL and MySQL) into every selected cell; Generate UUID / Generate UUIDv7 writes a different UUID into each one (uuid, text and long enough character columns). One staged change, one Undo.
- Copy: Copy as Markdown in the grid menu, and Settings > Results > Copy rows as chooses what Cmd+C copies: TSV (the default), TSV or CSV with headers, CSV, JSON or Markdown. A copied text value that starts like a formula gets an apostrophe in front, as in CSV exports; Neutralize Copied Formulas in Settings > Results turns it off.
- Invisible characters (control, zero-width, special spaces and direction controls) show as visible marks, with their code points in the tooltip. Copy, export and search use the real text. Show Invisible Characters in Settings > Results turns it off.
- Cell Value draws PostGIS geometry and geography and MySQL
GEOMETRYvalues, with the shape type, point count, SRID and Copy WKT. pgvectorvector,halfvecandsparsevecvalues show their dimension, norm, minimum, maximum and first values. A value over 256 KiB opens as plain text at once. - Tables without a primary key: a table tab of a PostgreSQL or SQLite table edits and deletes rows by their location (PostgreSQL
ctidchecked withxmin, SQLiterowidchecked with the row's values), so of two identical rows only the one you picked changes. A row changed since it was read is refused ("the row changed since it was read, reload the data"), the page is read again after such an edit, and these edits get no Data Rewind entry. MySQL/MariaDB tables and query results still need a primary key: a change without one is refused and names the table. - Column source: hover a column header to see
schema.table.column, plus "Masked by a connection rule" when a rule hides it. Query columns read from a base table on PostgreSQL, SQLite and MySQL/MariaDB (not views) and data viewer pages; expression columns, DuckDB and MongoDB show none. - Freeze Up to This Column (column header menu of a table tab) keeps the leading columns in place while you scroll sideways; Unfreeze releases them. Kept per table, by column name.
- Display As (header or cell menu of a binary column): size, hex, UUID, UUID with swapped time (MySQL
UUID_TO_BIN(u, 1)), MessagePack or Protobuf without a schema. Kept per table in table tabs, for the session in query results; Cell Value shows the decoded tree. Copy Value stays raw, Copy as Displayed copies the shown text, and editing opens the raw value. - Paste rows (table tabs; not DuckDB, MongoDB or read-only connections): Cmd+V, Edit > Paste or Paste Rows in the cell menu stages the clipboard's rows as inserts, one Undo for all, at most 10,000 rows. Fields split on tabs when the text has one, else on commas. A first row that names columns maps by name; otherwise columns follow the grid's on-screen order. An unquoted
NULLis SQL NULL; an empty field is''on text columns and NULL elsewhere. A value that does not fit its column stages nothing and names its row and column. While a cell is being edited, Cmd+V pastes into the cell. Known limits: a text cell holding the word NULL, copied with Copy Rows, pastes back as SQL NULL, and with copied formulas neutralized the leading apostrophe pastes back too. - Reload banner: when a SQL tab, a commit or an inline edit writes to a table open in another tab, that tab shows "This table changed in another tab." with Reload, which keeps its staged changes. Writes in a pending Protected-mode or your own transaction mark tabs at Commit and are dropped at Rollback. Writes from MCP, Shortcuts, Copy Objects, Data Rewind and dump restores do not mark tabs. After a data import, tabs on the written tables with no staged changes read their rows again instead (hidden ones when next shown); tabs with staged changes keep the banner, and an import in a Review transaction does this at Commit only.
- Large binary cells: a table tab of a PostgreSQL or SQLite table, or of a MySQL/MariaDB table with a primary key, reads binary values up to 64 KB. A longer one shows
<N bytes, not loaded>and is loaded by its key or row location when you open it in Cell Value or copy it. A value that was not loaded is never written: Duplicate Row and Copy as INSERT load it first or refuse. - Long text values: the same table tabs read text, JSON and XML values as their first 16,384 characters, shown with "…". Cell Value and Row Detail load the full value (JSON and XML are formatted only then), and copies and edits load it first. In editor and notebook results a value is cut at 1 MiB and a result stops at 256 MiB; the footer counts cut values, a saved notebook keeps them marked, and they are never copied, exported or edited (export All rows instead).
MongoDB
- Connect with a connection string or the form. Logins use SCRAM-SHA-256 or SCRAM-SHA-1 (a user created in
adminneeds Auth Sourceadmin). TLS is off or Verify Full, with an optional custom CA. - Commands: cells and SQL files take a subset of the mongosh syntax (
find,aggregate,countDocuments,distinct,insertOne/insertMany,updateOne/updateMany,replaceOne,deleteOne/deleteMany,show collections). Dblore parses it itself and never runs JavaScript. - Browse: click a collection in the sidebar. The filter bar takes a filter document such as
{status: "active"}and runs it on the server. - Grid or Documents: the picker under a result switches between the grid and document cards. A card shows the top-level fields; click a nested object or array to expand it.
- Edit in place: double-click a value on a card (or a grid cell) to edit it, Return to send, Esc to cancel. Each edit runs one
$seton that field after the Safe Mode confirmation. A double-click on an object, array or other type opens the field editor. The card header opens Row Details in the right sidebar, and a field's right-click menu has Show Details. - Add, duplicate, delete: in the data viewer, stage them by
_id, preview the exactinsertOne/updateOne/deleteOnecommands, then Commit. Results whose_idmixes types are read-only. - Build Aggregation… in the collection menu builds a pipeline stage by stage, with a read-only preview of each stage.
- Collection sections: next to the collection title, Documents, Aggregations, Schema and Indexes switch the view. Schema analyzes the sampled documents; Indexes lists each index with its type, size, usage and status.
- Filter bar tools: the options button sets projection, sort, collation, hint, skip, limit and
maxTimeMS; the bookmark button saves queries; Explain shows the plan as a tree; Export Code writes the query as Node.js, Python, Java, Go or C# driver code. - Aggregations: run a pipeline from the Aggregations section, built stage by stage or typed as text, with Wizard templates, saved pipelines, Explain and Export. Click a stage for the common operators, double-click to type any operator.
- AI button: the filter bar and the Aggregations section write a filter or pipeline from a prompt (Stop, then Apply).
- Administration: the Indexes section creates, hides and drops indexes, a Validation section edits the collection's validator, and the sidebar creates, renames and drops collections and views. The same commands (
createIndex,createCollection,collModand others) work in a.jstab. Drops ask for the typed name first. - Bulk Update / Delete… in the Documents header changes every document that matches the filter, after a count preview; an empty filter or 1,000 or more documents needs the collection name typed.
- Databases: the Databases button in the sidebar lists the server's databases; create, drop or switch to one (the workspace reconnects).
- JSON view: the picker under a result also offers JSON. The view mode and page size are remembered per collection.
- Export and import: Export... writes a collection, a query or an aggregation as Extended JSON or CSV; Import Documents... reads a JSON, NDJSON or CSV file after a preview.
- Back Up Database / Restore Backup (File menu) save collections and views, with indexes and validators, to one
.ndjsonfile and read it back. - An aggregation that ends in
$outor$mergeruns from the Aggregations section after the Safe Mode confirmation. - Query files: a new query in a MongoDB workspace is a
.jsfile;.sqlfiles are refused there. - MongoDB has no transactions, so writes ask before they run instead of waiting for Commit. For agents over MCP, use a user with a read-only role.
Structure editor
- Right-click a table in the sidebar and choose Edit Structure….
- PostgreSQL and MySQL: add, rename and drop columns; change type, default and nullable; create and drop indexes, with DESC per column (and NULLS FIRST or LAST on PostgreSQL); add and drop primary, foreign, unique and check constraints.
- SQLite: only the changes SQLite supports directly: add a column, rename a column (SQLite 3.25+), drop a column (3.35+), create and drop indexes. Constraints cannot be changed, and a key, indexed or generated column cannot be dropped; the table is never rebuilt.
- DuckDB: View Structure… is read-only.
- Changes are staged first. The preview shows exactly the SQL that runs, and Safe Mode confirms it like other schema changes. Each Apply is recorded in History.
- On PostgreSQL and SQLite the changes run together and join the pending transaction in Protected mode. On MySQL each statement commits on its own, so the editor is disabled in Protected (Review) mode, with pending changes or with your own transaction open; if a statement fails, it reports how many were applied.
- Closing the editor drops staged changes that were not applied. The editor and read-only object source tabs read the definition again when the window becomes active, after a schema refresh and when a Review transaction ends (at most once every 5 seconds; paused while one is pending). A clean editor updates in place; one with staged changes keeps them and shows "This table changed on the server since it was loaded" with Reload, which drops them.
- On the schema diagram: right-click a table header (Rename Table…, Add Column…, New Table…), a column (Rename Column…, Drop Column) or a foreign-key line (Drop Foreign Key). On PostgreSQL and MySQL/MariaDB, drag a column's grip onto a column of another table to add a foreign key; SQLite cannot add or drop a foreign key on an existing table. Changes collect in a bar with Preview SQL, Apply and Discard, and a renamed table keeps its place on the canvas. MySQL runs them one by one: a failure leaves the earlier ones applied.
- PostgreSQL types: New Type… in the Types section or Edit Type… on a type: add and rename enum labels (never drop), set a domain's default and NOT NULL and add or drop its CHECK constraints, add, drop, rename or retype composite attributes. Preview SQL and Safe Mode work as above. Before PostgreSQL 12 new enum labels are added on their own, outside a transaction, so they are refused while one is open or with the Review commit style.
- Table options (MySQL/MariaDB): the table's character set, collation and next AUTO_INCREMENT, applied in one
ALTER TABLE. Convert existing data rewrites every row withCONVERT TOafter a confirmation, and fails if some data does not fit. - File > Table Health… checks up to 2,000 tables on a separate connection: no primary key and foreign keys without a covering index (PostgreSQL, MySQL/MariaDB, SQLite), unused indexes and tables with many dead rows (PostgreSQL), unused and redundant indexes (MySQL
sysschema; MariaDB without it gets a note). Filter by Keys, Indexes or Bloat, jump to a table with Show Table, or Copy Report as Markdown. It refuses to run inside an open transaction.
Compare Schemas
- File > Compare Schemas… compares two schemas of the same engine: PostgreSQL, MySQL/MariaDB or SQLite. The target is a schema of this tab's connection, the only place a script ever runs. The source is another schema of this connection, or a saved connection of the same engine, read on a short-lived connection that is closed after the read.
- Compared: tables only in the source (create) or only in the target (drop), and in tables of both, columns added, dropped or changed (type, nullability, default), indexes and foreign keys added, dropped or changed. Views, routines, types, primary keys, unique and check constraints, comments, identity and generated columns are not compared. A renamed table, column or index shows as a drop plus an add.
- Tick the items to include. The script pane shows the exact SQL that Run sends, with Copy Script and Save Script….
- Drops of tables, columns, indexes and foreign keys are unticked by default; ticked drops ask first (Run with Drops), since their data cannot be recovered. Lossy type changes are flagged, and changes that cannot be scripted are listed to make by hand instead of run.
- Run goes through Safe Mode like the structure editor: on PostgreSQL and SQLite the script runs as one and joins the pending transaction in Protected mode; on MySQL each statement commits on its own, stopping at the first failure, with a report of what ran. A connection that changed since the compare is refused before anything is sent. Compare again after a run.
- A schema with more than 1,000 tables is refused, with its table count.
Copy Objects
- File > Copy Objects…, or Copy to Another Connection… on a table (ticked for you) or a schema. The source is this workspace's PostgreSQL, MySQL/MariaDB, SQLite or DuckDB connection; the target a saved PostgreSQL, MySQL/MariaDB or SQLite connection. Read-only and Schema-only targets are refused, and a locked target asks to unlock first.
- Tables only: columns, primary key, NOT NULL, literal defaults, unique constraints, plain indexes and foreign keys between the copied tables; CHECK constraints only between the same engine. Triggers are listed, never copied. A table name that already exists on the target refuses the copy.
- Types are mapped for the target, with a warning for each lossy one: PostgreSQL enums and arrays and DuckDB lists become text or JSON text,
timestamptzbecomesDATETIME(6)in UTC on MySQL, and unbounded text in a MySQL key becomesVARCHAR(255). - Preview SQL shows the planned statements, which you can edit (types, constraints, CHECK text). Renaming a planned table or column, or any other statement, refuses the copy.
- The target's Safe Mode asks before anything is sent (Review asks like Confirm). The copy commits on its own, outside the target's pending Protected-mode transaction, with progress and Cancel, and is recorded in the target's History.
- Rollback: on PostgreSQL and SQLite the whole copy is one transaction, so Cancel or a failure leaves nothing. MySQL/MariaDB commit each
CREATE TABLEat once: Dblore rolls back the rows, then drops only the tables this run created and names any it could not drop. ACREATE TABLEthe server committed just as Cancel cut off its reply is neither dropped nor named, so check the target after a cancelled MySQL/MariaDB copy. - Rows are read in chunks of at most 500 rows, so a copy is not a snapshot across tables.
Charts
- Switch any result between grid and chart with the toggle beside the result actions.
- Chart types: bar, line, area, point.
- Split by draws one small chart per value of a column (up to 12), with shared axes and one legend. The PNG export includes the whole grid.
Export
- Formats: Excel, CSV, TSV, JSON, Markdown, PDF and SQL
INSERT. - Copy results as SQL
INSERTor anINlist. - Per-format options with reset, redaction, wrapping and a size check.
- All rows: a result cut by the row cap, or one with cut values, can export every row as CSV, JSON or SQL
INSERT. The query runs again in a read-only transaction on a separate connection (DuckDB: the tab's own session; one read-only statement, no:nameparameters; not MongoDB), and masked columns stay masked. - File > Export SQL Dump… writes a PostgreSQL, MySQL/MariaDB or SQLite database to a
.sqlfile. Each table has Structure, Data and Drop; the database row sets them all (a dash when mixed). Drop writesDROP … IF EXISTSfor the tables, views, sequences, domains and types the dump creates, never CASCADE; on MySQL each drop commits at once. With Structure off for every table the dump holds only rows and restores into the database it came from, whose triggers then fire. - File > Restore SQL Dump… checks every statement of a Dblore dump before running it. It accepts drops only from a dump that wrote them and only for objects the same dump creates; on SQLite it refuses when a table outside the dump cascades from a dropped one. Older Dblore versions refuse dumps with drops. On PostgreSQL the dump runs with
search_path = pg_catalog, pg_tempand your value is put back afterwards, so a trigger or function the restore fires must use qualified names. - On MySQL/MariaDB connections the restore also reads mysqldump and mariadb-dump files: their session settings, per-table character set lines,
LOCK TABLES,DISABLE KEYS,DROP TABLE IF EXISTSfor tables the file creates, and rows of literal values.SET NAMES utf8runs asutf8mb4, and the session's settings are reset after the restore. A line Dblore does not restore is refused before anything runs, naming the option to dump with: GTID lines (--set-gtid-purged=OFF), triggers and routines (--skip-triggers, no--routines),USE(no--databases), views (--ignore-tablefor each view) and other character sets (--default-character-set=utf8mb4). The file must be valid UTF-8, so binary columns written as raw bytes cannot be read. - The restore has no file size limit: the file is read twice (checked, then run) and one statement can be up to 64 MiB. Reading and splitting the file show progress and can be cancelled, and a data-only restore fires the target's triggers (the dump and restore sheets say so). File > Copy Database File… saves an open SQLite database as one file. A dump holds your data in clear text.
- File > Export Data Dictionary… writes the columns, indexes and foreign keys of every table as a Markdown or a standalone HTML file (no scripts, no remote content).
Import data
- Import Data… from the File menu or a table's context menu: CSV, TSV, JSON and newline-delimited JSON.
- Preview rows and
NULLvalues, map columns, override inferred types, review the SQL, then import into an existing or a new table. Reading a large file shows progress, and a running import can be cancelled. - Limits: CSV, TSV and NDJSON files into a new table, or inserted into an existing one, stream with no file size or row limit. Keyed modes, JSON arrays and Import Files read up to 256 MB and 100,000 rows.
- Mode for an existing table: Insert, Update, Insert or Update, or Insert, skip existing. Rows match by the primary key or a unique index over plain columns, counted on the server (its collation decides); a key found more than once, or repeated in the file, refuses that row. A summary of new, existing and refused rows comes before the one Safe Mode confirmation.
- Import Files… (database, schema or table menu) imports several files, or the supported files of a folder (not its subfolders), in one transaction with one Safe Mode decision. Each file goes to the table with its name (case-insensitive) or a new table; untick a file, change its table or mode, or Edit its mapping. On MySQL/MariaDB creating tables commits implicitly, and the sheet says so.
- Generate Rows… (table menu; PostgreSQL, MySQL/MariaDB, SQLite) adds 1 to 1,000,000 rows of sample data from a seed: values by column type and name, a NULL ratio, foreign keys from up to 10,000 existing keys, and keys that never repeat within a run. Identity, generated and default columns are left to the server. A 10-row preview, then one transaction with progress; Cancel rolls it back. In Protected mode the rows wait for Commit.
Safe Mode
- Five protection levels.
- Confirmation before
DROP,TRUNCATE,ALTERandCREATE. Warning onDELETEorUPDATEwithoutWHERE. - Pending transactions show a Commit / Rollback banner; View Details shows every pending statement in full. Touch ID, or your Safe Mode password, unlocks changes.
- Commit style (footer or Settings > Security): Immediate, Confirm, Review or Password, for dialogs, unlocks and inline edits.
- Read-only connections block writes. Unlock writes (1 or 5 minutes) in the protection menu of a connected Read-only connection lets data writes run after the Safe Mode authentication, with a countdown and Lock now; schema changes stay blocked. It is never saved, ends on expiry, Lock now, relock, disconnect or a session reset, and MCP, the command line and Shortcuts never get it. Read-only DuckDB connections and SQLite files opened read-only cannot unlock (their file is open read-only).
- Related rows before a delete: when staged changes delete rows, Preview SQL and the Safe Mode dialog list what else the delete touches: rows deleted by cascade, rows set to NULL or to their default, and rows that still refer to the deleted ones. A foreign key that would block the delete disables Commit and names the constraint; Commit Anyway lets the database decide. The counts run on a separate read-only connection within 5 seconds, and are skipped while a transaction is open.
- Large reads inside transactions are capped with a server-side cursor (PostgreSQL).
- MySQL commits implicitly before DDL,
TRUNCATE,LOCK TABLES, grants and similar statements, so Protected mode refuses them. Turn Protected mode off for the connection to run them.
Lock
- Lock this connection (Safety card of the connection form) locks a saved login (every database of that host, port and user) or a SQLite or DuckDB file. Connecting, testing, editing, removing or exporting it asks first: Touch ID or the Mac password, or the Lock password when one is set. Files are guarded, not encrypted. A saved password shows only after Touch ID or the Mac password.
- File > Lock Workspace… encrypts the
.sqlwsand every tab file it saves (notebooks,.sqland.jsscripts, Markdown notes), plus the unsaved text kept for the next launch, with AES-256-GCM. It needs a Lock password of 8 or more characters: without it or this Mac's Keychain, the files cannot be opened. Remove Lock… writes them back as plain files. - The lock button beside Details in the title bar of a saved workspace opens Lock Workspace; on a locked workspace it locks and closes it at once, and the next open asks again.
- A locked workspace (Welcome, Finder, Open Recent or launch restore) opens its window first, then asks; content shows after unlock. Recents keep its name with a lock mark. When the Keychain cannot keep the key behind Touch ID, the workspace opens with the Lock password only, and Dblore says so.
- Dblore relocks on sleep, screen lock, idle time (15 minutes by default) and quit: windows of locked items show a cover with Unlock, most of their menu commands are off and running AI requests stop. Connections stay open.
- Settings > Security > Lock sets, changes or removes the Lock password, chooses the relock time and lists locked items. Changing the password re-encrypts the locked workspaces in Recents at once and lists any it could not reach; each keeps its key, so older copies and backups still open with the old password.
- Settings > Data asks to unlock before it clears history, imports a backup or deletes Lock secrets while something is locked. Deleting all secrets removes the Lock password and the workspace keys: locked workspaces then open only with the Lock password.
- The MCP server never serves a locked connection.
- Older Dblore versions refuse locked workspace files and never overwrite them, but drop connection locks when they save the connection list: lock those connections again afterwards.
AI assistant
- Toggle the chat panel with Cmd+L. It writes SQL from plain-language questions and explains or fixes queries.
- Bubble mode: the detach button in the panel header turns it into a floating circle at the bottom of the window. Click the circle to expand or collapse the chat, drag it left or right to move it, hover it and click the small x to hide it, and press Esc to collapse it. Choose the default in Settings > AI > Panel (Right sidebar or Bubble).
- Providers: Anthropic, OpenAI, Google Gemini, OpenRouter, xAI, Requesty (its router forwards requests to the model vendor), ChatGPT sign-in, Ollama, LM Studio,
mlx_lm.server, any custom OpenAI-compatible endpoint, Apple Intelligence, or on-device MLX models (Qwen3.5 2B, 4B and 9B, Gemma 4 12B). Settings > AI > Unused model files removes files of models that are no longer offered. - On-device models need the On-device AI plugin: without it, Settings > AI shows Install…. Models download separately from Hugging Face; every file is checked by its size and SHA-256 before the model is installed, and models are kept when you remove the plugin.
- Apple Intelligence (macOS 26 or later) uses Apple's model on this Mac, with no key; the provider shows dimmed with the reason when this Mac cannot run it. It has a small context (4,096 tokens before macOS 27), so Dblore trims the schema and history to fit, and tools are off as with other on-device models. Chat, Review, Explain, Fix and inline suggestions use it. Older Dblore versions drop these AI settings.
- Bring your own API key. Claude Pro/Max subscriptions are not supported.
- Sent automatically: table and column names, types and keys.
- Sent only after you review and press Send: query text and error messages.
- Row data is never sent on its own. Generated SQL does not run automatically.
- Run read-only queries (Settings > AI > Privacy, off by default): the assistant writes a query, Dblore runs it and shows you the rows. Only a single read-only statement runs, in a read-only transaction; anything else is refused and shown as SQL. Row values and errors are never sent to the model.
- In bubble mode, drag an edge of the card to resize it.
- The chat can also call tools of MCP servers you add, one approved call at a time. See Outside MCP servers.
- Change settings: ask the chat to change Dblore's settings (theme, editor, formatter, results, notifications…). The change applies at once and shows a card "old → new" with Undo. It works with no connection and on connections whose AI access is off, since it reads no connection data, but not with on-device models. Security, AI provider and keys, MCP, Shortcuts, Safe Mode and session limits stay manual: the assistant names the Settings tab instead.
/settingsfills the message box with a settings request, and the empty chat panel suggests examples. - For fully local use, point it at Ollama, LM Studio or
mlx_lm.server.
MCP server
- Lets AI agents such as Claude Code and Cursor read your databases through the Model Context Protocol. Off by default: turn on Run the MCP server in Settings > MCP. Dblore must be running.
- The server listens on this Mac only, at
http://127.0.0.1:<port>/mcp(default port 41730). Use only the port Settings shows as running. - Every request needs the Bearer token from Settings > MCP. Copy Claude Code command copies a
claude mcp addcommand to paste in a terminal; Copy Cursor JSON copies the entry for Cursor'smcp.json. Regenerate creates a new token; clients with the old one stop working until you copy the setup again. - Agents see only the connections you turn on in the list, never a locked one, even after you unlock it in Dblore. A connection that asks for its password on connect is refused. DuckDB connections cannot be shared.
- Paired command-line tools use their own tokens and read only. See Command-line tool.
- At most 16 open connections and 8 requests at a time; above that the server answers HTTP 503 with
Retry-After: 2. - Tools:
list_connections,list_tables,describe_tableandrun_query, plusexecute_writewhen write tools are allowed (see MCP write tools). run_queryis read-only: one statement per query, checked first and then run in a read-only transaction (SQLite files open read-only). Writes, DDL,SET,CALLand several statements are refused. A query returns 200 rows by default (at most 1000), cells are cut at 2000 characters and a query stops after 30 seconds. A timestamp without a time zone comes as its stored wall clock (2026-10-15 12:00:00.123), a zoned one as ISO 8601 UTC with its fraction (2026-10-15T12:00:00.123Z); a NUMERIC or DECIMAL value is a JSON number when it has at most 15 significant digits, else its exact text.- Use a least-privilege database user: an agent can read everything the connection's role can read, including other databases on the same server.
- Query results go to the agent. Treat them as untrusted data: text stored in a table can try to give the agent instructions.
- Agent queries are recorded in History (when History is on). They do not see uncommitted changes in Dblore's own open transaction.
MCP write tools
- Off by default: turn on Allow write tools in Settings > MCP. Agents can then ask to run one change at a time through the
execute_writetool. - Once write tools are on, each connection turned on in the list gets an Agent access scope: Read (the default, no changes), Read + data changes (
INSERT,UPDATE,DELETEandMERGE; MongoDB inserts, updates and deletes) or Read + data + schema changes. - Schema changes cover tables, indexes, views, sequences, schemas, types, domains and MySQL databases,
TRUNCATE,RENAME TABLEandCOMMENT ON; on MongoDB,createCollection,createIndex,dropIndex,dropandrenameCollectionin the same database. - Refused before anything is asked: creating, changing or dropping functions, procedures or triggers (use the SQL editor); on MongoDB,
collMod, hiding or unhiding an index,createView,dropDatabase, system collections and pipelines that name another database; statements with hidden characters, privilege or account changes, extensions, foreign data or file access; more than one statement; aDELETEorUPDATEwithoutWHEREwhose single target table cannot be read (multi-table forms,WITH … DELETE,DELETE … USING). - Changes run with the connection's protection and commit style. A connection whose commit style is Review or Password refuses agent changes: set Confirm or Immediate to allow them.
- Every call opens the MCP Write Request panel, even when only the welcome window is open. It shows the connection and engine, the kind of change, the exact statement and the client name (as the client reports it, unverified). Approve once runs it once; Deny or Deny all refuse. A destructive change (
DROP,TRUNCATE,CREATE OR REPLACEof a table, schema or sequence, a dropped column,MERGEthat deletes,DELETEorUPDATEwithoutWHERE; MongoDBdrop()anddeleteMany({})) needs its target's name typed before Approve works. Return never approves, and Approve is ignored for a moment after a request shows, so a click meant for something else never approves it. - There are no standing grants: every call asks again. Requests show one at a time, at most three wait (a further one is refused at once), and a request not answered 2 minutes after it shows is denied. Turning off write tools or MCP, or taking a connection's scope away, denies the waiting requests.
- Approve only statements you have read: an agent can be steered by text stored in the database.
Command-line tool
dbloreships inside the app. Settings > MCP > Command-line tool copies the line that links it into/usr/local/bin(it runssudo, so you paste it in a terminal yourself).- It needs Dblore running with its MCP server on. Pair once: Pair Command-Line Tool shows
dblore pair --port <port>and a code; run the command and type the code (not echoed), orpbpaste | dblore pair --port <port>. A code works once, for 5 minutes, and is burned after 5 wrong tries. dblore connectionslists the connections it may use: those shared with the MCP server, except connections whose AI access is Never. Locked, Vault, password-command and Ask on connect connections are refused, as for MCP.dblore query --connection NAME "SELECT …"(or-f file.sql) runs one read-only statement, under the same rules asrun_query.--format csv,tsv,json,jsonlortable(the default) and--max-rows N(up to 1,000). Rows go to stdout; truncation notes and the untrusted-data notice go to stderr.- Exit codes: 0 done, 1 usage, 2 Dblore is not running or its MCP server is off or busy, 3 refused (pair again after a revoke), 4 query error.
- Before it sends its token, the tool checks Dblore's proof for the port it dialed, so another program on that port never gets the token. Its pairing is kept in
~/Library/Application Support/Dblore CLI/credentials.json, readable by you only. - Settings lists paired tools with Revoke, which applies at once. Queries are recorded in History as Command line; older Dblore versions show these rows as editor queries.
Shortcuts
- Dblore adds Open Connection, Run Query and Add Row actions to the Shortcuts app. They work with the MCP server off.
- Off by default: turn on Allow Shortcuts in Settings > Shortcuts, then the connections Shortcuts may use (its own list). What Shortcuts refuses lists the rest.
- Refused: locked connections, connections that would need you (an SSH host key not trusted yet, a secret not in the Keychain), Vault connections, connections with a password command and Ask on connect connections; Open Connection also refuses one with a pre-connect or post-disconnect script. Shortcuts never runs your scripts.
- Run Query runs one read-only statement and returns at most 1,000 rows (1 MB) as CSV. Add Row takes a table and a JSON object of values and inserts one row, only with the Immediate commit style and Protected mode off, since Shortcuts cannot ask you.
- Actions are recorded in History as Shortcuts; older Dblore versions show these rows as editor queries. Query results are untrusted data.
Outside MCP servers
- Settings > MCP > Outside MCP servers adds an MCP server whose tools the AI chat may call: a name, a URL and an optional token.
- The URL must be
https, or plainhttpto127.0.0.1,::1orlocalhost. Redirects are never followed. - The token is sent as a Bearer token and kept in the Keychain. It is listed and deleted with the other secrets in Settings > Data.
- A new server starts turned off, with no tools allowed. Test connects to it and lists its tools; tick the ones the AI chat may use, then turn the server on.
- If the server later changes a tool's description or schema, the tool shows Changed and is unticked until you review it and tick it again. Text from the server is shown with hidden and direction characters removed.
- In the AI chat: on a connection whose AI access allows it, the chat can offer the allowed tools of enabled servers, never with an on-device model. Each call shows a card with the server, the tool and the exact arguments. Nothing is sent until you press Run once, which accepts clicks only after the card has been shown for a moment; every call asks again. Deny tells the assistant the call was denied.
- Results are untrusted: the tool's result goes back to the assistant marked as data from a third party, never as instructions, and nothing in it can run SQL. Arguments the card cannot show exactly (hidden characters, or longer than the card) are refused. A server that does not answer within 10 seconds is left out of that turn, with a note in the chat.
- Add only servers you trust: a server can return text that tries to steer the AI.
Plugins
- Optional engines and features, installed from Settings > Plugins. Dblore downloads a signed, notarized library or bundle and checks its SHA-256 and signature before it loads it. Remove deletes it; a plugin in use stays loaded until you quit Dblore.
- DuckDB: open a
.duckdbfile (read-only by default), create one or use an in-memory database. Query Parquet/CSV File… in the command palette queries a file directly. - Safe Mode guards DuckDB writes such as
INSTALL,ATTACH,COPY … TOandCREATE TABLE. Data import works, except in Protected mode or with a transaction open (DuckDB has no savepoint); row edits, Paste Rows, Generate Rows and foreign-key lookup are not available. - SQL Server: browse schemas, tables, views, procedures and functions, page and filter tables, edit rows of a table with a primary key, follow foreign keys and Stop a statement. Not available: the structure editor, SQL dump, import, Explain, Data Rewind, Find a Value in Every Table, Table Health, Generate Rows, Schema Compare, Copy Objects, MCP and Shortcuts. A
GOline is refused; the editor splits statements at;, so a T-SQL variable does not reach the next statement, and a procedure that returns several result sets shows only the last one. Built on FreeTDS (LGPL-2.1) and OpenSSL (Apache-2.0); Settings > Plugins links their licenses, the plugin source and the exact FreeTDS source. - ClickHouse: browse databases as schemas, query, page and filter tables, Stop a query and Explain as text. Not available: grid edits, transactions, the structure editor, SQL dump, import, Data Rewind, Find a Value in Every Table, Table Health, Generate Rows, Schema Compare, Copy Objects, MCP and Shortcuts.
- Redis: a Redis workspace opens command tabs (
.redisfiles) instead of notebooks. Run the current line, the selection or every line in a database picked from db0 to db15, see each reply as Grid or Raw, and Cancel a running command; a reply stops at 10,000 rows. The sidebar lists db0 to db15 with their key counts and each database's keys in folders split at:, filtered by a match pattern (up to 10,000 keys per database; bytes that are not printable text show escaped). Click a key to open it in a tab: type, TTL and size, then a string as Text, Hex or JSON (the first 256 KiB), or a hash, list, set, sorted set or stream in pages of 200. Edit a string, a hash field, a list element, a sorted set score or the TTL; add fields, elements, members and stream entries; remove fields, members and stream entries; rename or delete the key; create one with New Key in a database's context menu. Removals always ask first, other edits follow the commit style, and every edit is listed in History. Key tabs are not saved with the workspace. - Redis safety: Dblore checks every command before it is sent. Read-only allows reads only; Schema Protected allows data writes but not admin, destructive, script or unknown commands.
FLUSHALL,FLUSHDB,KEYS,CONFIGand other destructive commands always ask first. Command tabs refuse commands that change the shared connection (SELECT,AUTH,MULTI), subscribe or wait (SUBSCRIBE,BLPOP),SHUTDOWNandDEBUG. Read-only is checked by Dblore and the plugin, not the server, and means no keyspace writes:EVAL_RO,EVALSHA_ROandFCALL_ROstill run, and Cancel or the statement timeout can sendSCRIPT KILLorFUNCTION KILLto stop a running script, also on a read-only connection. For a guarantee on the server, use an ACL user without write commands. A command that carries a password (such asCONFIG SET requirepass) is not written to History. Not available: notebooks, the schema visualizer, the AI assistant, result export, MCP and Shortcuts. - On-device AI: runs AI models on this Mac with MLX for the AI chat and inline suggestions. See AI assistant for models.
Settings & data
- Open with Cmd+,. View keyboard shortcuts in Settings > Keyboard.
- Search sections from the sidebar of the settings modal.
- Light or dark theme (also the toggle in the header) and an accent color. Settings > Appearance > Animations turns interface animations on or off; System follows macOS Reduce Motion.
- When queries, imports, restores, dumps, copies, exports or schema compares are running, Quit (also from the Dock or the menu bar), closing a workspace window and Disconnect list them first (a Parquet export is listed at Quit only): Stop stops them and goes on, Cancel keeps them running. A restore says whether stopping rolls it back or keeps what already ran (MySQL, MongoDB); Stop waits for a running restore's own cancel warning, and "Keep running" cancels the quit, close or disconnect.
- Settings > General > Confirm before quitting with Cmd+Q asks first when nothing runs. Quitting from the Dock or the menu bar then never asks.
- Long-query notifications (Settings > General, off by default): a macOS notification when a query runs longer than a threshold. It never shows the SQL or data; clicking it opens the tab.
- Editor font size, word wrap, side-by-side layout and a configurable result row cap.
- Safe Mode levels. AI keys are stored in the Keychain.
- Updates: toggle background checks.
- Data tab: export, import or clear local data as a
.dblorebackuppackage, with progress, a Cancel button and a toast when it ends. Secrets are never included. Saved SSH passwords, keys, trusted host keys, command-line pairings and Vault secrets can be deleted here.