Back to search
agentable
SkillGitHub
Native macOS database client for people and their coding agents
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/subasioguz--agentable)HTML
<a href="/listings/subasioguz--agentable"><img src="/api/listings/subasioguz--agentable/badge.svg" alt="findsafeskills safety rating" /></a>Install
i
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 190 lines
- <p align="center">
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="docs/brand/wordmark-dark.svg">
- <img src="docs/brand/wordmark.svg" alt="Agentable" width="320">
- </picture>
- </p>
- <p align="center">A native macOS database client for you and your coding agents.</p>
- 
- You browse, filter and edit data in the app. Your agents (Claude Code, Codex, Cursor and others)
- work with the same saved connections through a CLI or an MCP server. They never see your
- passwords, they only do what each connection allows, their changes to protected databases wait for
- your approval, and everything they run is logged.
- > Status: early development. The app, the CLI and the MCP server work with SQLite and PostgreSQL.
- > There is no signed download yet; build it from source (see Build below).
- > See [docs/PLAN.md](docs/PLAN.md) for what comes next.
- ## The app
- | Filters with value counts | Review & apply |
- |---|---|
- |  |  |
- | **Advanced conditions** | **Dark theme** |
- |  |  |
- - **Connections**, grouped by environment. Paste a `postgres://` URL or pick a SQLite file; the
- password goes to Keychain. An SSH tunnel runs through your own OpenSSH, so `~/.ssh/config`,
- `ProxyJump`, your SSH agent and `known_hosts` work as in Terminal (keys only, no SSH passwords;
- unknown host keys are refused). TLS can check a CA file and send a client certificate.
- - **Schema rail** with row counts, and columns with their types, keys and references.
- - **The table**, in pages: the bar under it says which rows you see ("1,201–1,400 of ~1.2M rows"),
- ‹ › (⌘[ ⌘]) turn pages, and the page button jumps to a page, sets rows per page (10 to 1,000, or any
- number you type) and turns an estimated total into an exact count. Filters and sorts start again at page 1. Each table remembers its
- sort, column order, hidden columns and widths; "Reset layout" in Columns puts it back, and
- Settings picks whether tables open oldest or newest first.
- - **Search rows** across every column with ⌘F, without making a filter; **Go to anything** (⌘K)
- jumps to a table or another connection.
- - **Filters** as chips with a value picker that shows how many rows have each value. "Advanced"
- turns the same filters into condition rows with and/or groups, and "WHERE" takes plain SQL.
- - **Your edits wait.** Double-click a cell to change it, add a row with ⌘I or duplicate one.
- Changes wait in a list under the table, where you can undo any one of them (⌘Z steps back too).
- ⌘S saves them in one transaction; on production you see the SQL first. When agents have proposals
- waiting, your edits and their proposals go through Review & apply together. An edit is refused if
- its row is gone or the value you edited changed since you loaded it (decimals and timestamps are
- not compared, since they cannot be matched exactly).
- - **Row details and foreign keys.** Space opens the selected row at the side: long text wraps,
- JSON is indented, and a foreign key opens the row it points to, with a way back. "Referenced by"
- lists the tables that point at the row and how many of their rows do.
- - **SQL tab** next to Data and Structure: colored SQL with completion of tables, columns and
- keywords. ⌘↩ runs the statement under the cursor; reads run on your read-only session, writes
- get a trial run and wait with your other changes. Explain shows the plan. The side panel lists
- saved queries (`.sql` files) and what ran on the connection lately, by you and by agents.
- When an agent suggests a new version of a saved query, its lines show in the editor, tinted,
- with its note: Tab accepts (⌘Z undoes), Esc dismisses; nothing runs until you run it.
- **Ask** (⌘I) sends a request about the open query ("only paid orders", "why is it slow?") to
- your Claude Code, which works in the background with Agentable's tools only: a change comes back
- as a suggestion in the editor, an answer as text above it. See Privacy below.
- - **Structure.** The Structure tab lists columns, indexes and checks (on PostgreSQL also index
- sizes, how often they are used and the views that read each column). Add, rename or drop a
- column, change its type, default or whether it allows empty values, and add or drop indexes.
- Changes wait as DDL with your other edits and run in one transaction; on production you see each
- statement and what it locks first. SQLite only renames a saved column in place. A renamed column
- keeps its mask for agents. An agent's proposed index shows in the list with its reason.
- - **Copy and export.** Copy rows as CSV, JSON, SQL INSERT or Markdown, or save the selection, the
- filtered view or the whole table to a file.
- - **Agents in the same table.** Rows an agent wants to change are tinted orange with the new value
- in a dashed box, its cursor shows the column it last read, and the Session panel lists what each
- agent did today.
- - **First launch** offers to add Agentable's MCP server to Claude Code, Codex and Cursor; later,
- "Connect agents" on the Connections screen or in the settings menu does the same.
- - Light and dark themes. The interface is in English, German, Spanish, French, Italian, Brazilian
- Portuguese, Russian, Turkish, Japanese, Korean, and Simplified and Traditional Chinese; it follows
- your Mac's language, or pick one under Language in the settings menu.
- ## Connect your agents
- `agentable init` prints the setup for Claude Code, Codex and Cursor. Agents then get eight MCP tools:
- `connections`, `schema_search`, `describe`, `query`, `explain`, `propose_change`, `saved_queries`
- and `suggest_sql`. The last two let an agent read your saved queries on a connection it may use
- (not ones that mention a hidden table) and suggest a new version of one, or a new query, which
- shows in your SQL editor until you accept or dismiss it.
- ## How agents use it from a shell
- ```sh
- agentable tables shop
- agentable search shop email
- agentable describe shop customers
- agentable query shop "SELECT name, email FROM customers LIMIT 20"
- agentable propose shop "UPDATE orders SET status = 'refund_pending' WHERE id IN (1, 2)" --reason "charged twice"
- agentable queries shop charged-twice.sql
- agentable suggest shop charged-twice.sql "SELECT ..." --note "count only successful payments"
- ```
- Results come back as compact tab-separated text (or `--json`), marked as data rather than
- instructions. A refused request exits with status 2 and says why.
- ## What each connection allows
- | Access | Agents can |
- |---------|-------------------------------------------------------------------------|
- | off | not see the connection |
- | read | read schema and run read-only queries within row and time limits |
- | propose | read, and file changes that a person reviews with `agentable review` |
- | write | read and write directly (never on production) |
- Defaults follow the environment: production starts at read, staging at propose, local and dev at
- write. Tables can be hidden and columns masked per connection. Raising access, unhiding or unmasking,
- and approving a change all ask for Touch ID or your password, so an agent with a shell cannot do them.
- ## Your side
- ```sh
- agentable connections add shop ./shop.sqlite --env production
- agentable connections rules shop --hide api_keys --mask customers.email
- agentable connections agent-role acme # SQL for a SELECT-only agent role (PostgreSQL)
- agentable proposals
- agentable review p-1a2b3c --approve
- agentable log shop
- ```
- ## Privacy
- Agentable has no account, no telemetry, no analytics and no crash reporting. It connects only to
- the databases you add. Fonts ship inside the app, so it never fetches anything to draw itself.
- One exception, only when you use it: **Ask** in the SQL tab runs your own Claude Code on this Mac,
- in an empty temporary folder, with nothing but Agentable's MCP tools (no files, no shell, no other
- servers) and no permission to propose changes. Claude Code then sends your request, the saved
- query and what it reads through Agentable to Anthropic, as it always does, under your Claude
- account. Nothing starts until you press Send, and Stop or quitting the app ends it.
- What it keeps, all on your Mac:
- | What | Where |
- |------|-------|
- | Saved connections (no passwords) | `~/Library/Application Support/Agentable/connections.json` |
- | Passwords | your login Keychain |
- | Audit log: who ran which statement, when, how many rows (not the rows themselves), including what you run in the SQL tab | `~/Library/Application Support/Agentable/state.sqlite` |
- | Proposals, with up to 200 of the rows they would change so you can review them (masked columns stay masked) | the same `state.sqlite` |
- | Agents' suggestions for your saved queries: the query's text when they suggested, their text and note | the same `state.sqlite` |
- | Window preferences: theme, default order, each table's sort, column order, hidden columns and widths, last opened table, panels you opened or closed, the last export format, when you last used each connection | macOS user defaults |
- | Saved queries | `~/Library/Application Support/Agentable/queries/<connection>/*.sql` |
- | Exports | only the file you pick in the save dialog |
- `AGENTABLE_HOME` moves connections, the state database and saved queries somewhere else.
- **What agents read leaves your Mac through the agent.** Query results go to the agent you
- connected, and the agent sends them to its model provider. Agentable decides what the agent gets:
- hide tables, mask columns, give agents a SELECT-only database role, and keep production at read or
- propose. Those rules apply to every agent request, and every request is logged, including refused
- ones.
- **Setting up agents** (first launch) changes their configuration only when you click Connect:
- Claude Code and Codex through their own `mcp add` commands, Cursor by adding one entry to
- `~/.cursor/mcp.json` after saving a copy of the file next to it. Before that, the app only reads
- those files to show what is already connected.
- ## Security
- The safety model is in [docs/PLAN.md](docs/PLAN.md#safety-model). To report a vulnerability,
- see [SECURITY.md](SECURITY.md).
- ## Build
- Requires macOS 15 and Swift 6.
- ```sh
- swift build
- swift test
- .build/debug/agentable --help
- scripts/bundle-app.sh # .build/Agentable.app with the CLI and fonts inside
- scripts/snapshot.sh # renders every screen to PNG with a demo database
- ```
- `scripts/snapshot.sh` keeps everything in its output folder (`AGENTABLE_HOME` points there), so it
- never touches your own connections.
- ## Brand
- The mark, app icon, colors and voice are in [docs/BRAND.md](docs/BRAND.md) and `docs/brand/`.
- ## License
- Apache License 2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE). The bundled fonts, Instrument Sans
- and Spline Sans Mono, are under the SIL Open Font License 1.1; their licenses are next to them in
- `Sources/AgentableApp/Resources/Fonts`.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.