Network controls
The remaining sections describe Mac features unless a platform is named.
Usage data
From Mac 0.86.0, including 0.86.0-beta.1, PostHog’s SDK sends directly tohttps://eu.i.posthog.com. It also requests /config from https://eu-assets.i.posthog.com and /flags from the ingestion host. Share usage data in Settings > General > Privacy is on by default. The first event needs no prompt. Starting with the switch off does not initialize the SDK. A build without a PostHog project key sends nothing.
The SDK adds model, screen size, time zone, language, Wi-Fi state, app version/build/bundle ID, OS and SDK versions, and session ID. PostHog receives the request’s IP and derives country/city. Events use a random installation ID before sign-in, without a person profile. Sign-in identifies the account and sets
app_version, os_version, architecture and locale on its profile; PostHog also adds derived location. Sign-out resets the random ID. No machine ID, license status or update settings enter these events.
No hostnames, usernames, SQL, rows or credentials enter usage data. The switch stops app events, not device registration or server events about sign-ins, purchases, devices and teams. It does not delete earlier records. See analytics and rights.
Linux usage data
From 0.1.0-beta.1, turn off Share Usage Data in Preferences > General > Privacy to stop new sends. It is on by default. A request already in flight can finish. Builds without a PostHog project key send nothing. Theposthog-rs client sends to https://eu.i.posthog.com. It captures the events in the Mac table except feature_gate_shown. signed_in.method is interactive; support_ask_answered.answer is buy. The license ask’s count of days used stays local. The SDK supplies library/OS metadata. Account identification sends $identify, $anon_distinct_id and $set with app_version, os_version, architecture and locale. Before identification, $process_person_profile is false. Sign-out resets the random ID. PostHog receives the IP and derives rough location.
The installation ID and first-event flags live in telemetry.json under the app’s state directory. No hostname, SQL, rows or credentials are event properties. Account device registration separately sends the hostname and machine ID. Turning usage data off does not stop account requests or server analytics. The Linux policy states retention and rights.
Older Mac reports
Before 0.86.0, the default-on switch controls aPOST to https://api.tablepro.app/v1/analytics, ten seconds after activation and every 24 hours while running. A 20-hour cooldown prevents another report after reopening.
X-Signature contains an HMAC-SHA256 of the JSON. The old Mac path does not report the first query. The current server relays the report to PostHog with Cloudflare’s country, without an IP address or license key, and without creating a person profile. The machine ID can link reports to account devices. The server no longer keeps a dedicated heartbeat history; processing jobs, logs, backups and the migration archive are separate. Before the November 4, 2026 server cutover, heartbeat records included IP addresses. See retention.
iPhone and iPad usage data
Nothing is sent until you choose Share Usage Data at onboarding or turn it on in Settings > Privacy. The app then sendsapp_opened, connection_attempted, connection_succeeded and first_query_executed to PostHog at https://eu.i.posthog.com, with a random installation ID and no person profile. iPhone and iPad lists their properties and what posthog-ios adds. The app makes no account, license, update or plugin check. See iOS privacy for the separate Apple diagnostics route.
Earlier versions sent a daily report to the relay endpoint above instead, also only after you chose Share Usage Data. Its fields were machine_id (SHA-256 of Apple’s identifier for vendor), platform (ios), app_version, os_version, architecture, locale, database_types, connection_count, has_license (always false), connection_attempted_at, connection_succeeded_at and first_query_executed_at.
AI providers
Nothing reaches a model until you add a provider, and a fresh install has none: with none configured, every AI path returns before it builds a request. Enable AI Features in Settings > AI is on out of the box, but it gates an empty tab. One provider is marked active and answers inline suggestions and the editor actions; the chat panel’s own model picker can send a turn to any other provider you configured. Requests go to that provider’s endpoint. The presets areapi.anthropic.com, api.openai.com, generativelanguage.googleapis.com, api.x.ai, openrouter.ai, router.requesty.ai, api.meta.ai, api.cursor.com and opencode.ai, plus http://localhost:11434 for Ollama and http://localhost:8080 for llama.cpp and MLX, which keeps their traffic on the machine. Every endpoint field is editable, and a custom provider goes wherever you point it. Meta’s -contributor models, which Meta may train on what you send, stay out of its model list; one is used only when you type its id under Other….
A chat turn carries:
- The database type and the name of the database you are connected to.
- The schema, while Include database schema is on (the default): table names, estimated row counts, column names and types, primary key and NOT NULL flags, DEFAULT values, and foreign keys, for up to Max schema tables tables (20 by default). With more tables than that, the ones your question and current query use come first, and the rest are sent as a count, not by name.
- The query in the current tab, while Include current query is on (the default), truncated at 2,000 characters.
- Result rows, only if you turn on Include query results, which is off by default. It attaches the first 10 rows of the current result, each value cut at 200 characters.
- The connection’s AI Rules text, if you wrote any.
- Any images you paste into the composer, and the conversation so far.
execute_query itself, and the rows it gets back become part of the conversation, so they go to the provider on the next turn. Those results are capped by the row limits in Settings > MCP: 500 rows by default, 10,000 at most.
Per connection, the AI Policy picker in the connection form’s Options section takes Use Default, Always Allow, Ask Each Time, or Never. The app-wide default is Ask Each Time, which asks once per connection per chat session before the first send. Never blocks the chat panel and external AI tool calls against that connection.
Provider API keys live in the keychain, never in connections.json and never in a CloudKit record. The provider list itself (type, name, endpoint, model) rides along with the Settings category of iCloud sync when sync is on.
Signing in to ChatGPT opens a second local listener on port 1455 for the OAuth redirect. It runs only during that sign-in and closes after it.
Two provider paths run a program rather than call an endpoint. Claude Agent runs the claude command line tool, which talks to Anthropic under your own subscription. GitHub Copilot downloads its language server from registry.npmjs.org and starts it on first use or when Settings > AI lists the provider, never at launch. It stops when AI is turned off or the Copilot provider is removed. Its provider sheet has Send telemetry to GitHub, which is on when you add the provider. Turning it off sets the language server’s telemetry level to off.
iCloud sync
Off by default, and it needs a Pro or Team license. When it is on, records go to the CloudKit containeriCloud.com.TablePro, into the private database of your own iCloud account, in the Production environment.
Synced record types are Connection, ConnectionGroup, ConnectionTag, AppSettings, SQLFavorite, SQLFavoriteFolder, FavoriteTable, FavoriteDatabase, SSHProfile, CredentialProfile, TableFolder and TableFolderItem. Credential profiles carry the name, username and password mode, never the password or command source. A connection record carries host, port, username, database, and SSH and SSL settings. Sync runs on launch, about two seconds after you change something synced, when you switch back to the app, and when iCloud notifies the app that another device changed something. That notice goes through Apple’s push service, carries no record content, and asks for no notification permission. The app registers for it only while sync is on.
Passwords are not one of those record types. The Passwords toggle nested under Connections instead marks keychain items as synchronizable, so Apple’s iCloud Keychain carries them end to end encrypted. It is off by default, and it applies to items written after you turn it on.
With password sync off, secrets are written kSecAttrAccessibleAfterFirstUnlockThisDeviceOnly, which keeps them on that one Mac. Query history, the activity log, the execution log, tab state, MCP settings, and the sync settings themselves are never synced. Individual connections can be marked Local only, which also moves their secrets out of iCloud Keychain onto that one Mac. Full detail in iCloud Sync.
The MCP server
Direct HTTP clients connect through IPv4 loopback. The bundled bridge and CLI use a private Unix socket that accepts only the same macOS user. Both listeners require authentication for data access. Neither has a remote mode. See MCP transport and handshake for paths and permissions. It is off by default, and nothing starts it while Enable MCP Server is off: not the bundled bridge, not the pairing flow, not atablepro://integrations/start-mcp link. What a paired client can reach is real data, so four gates stack:
- A token on every request. One that arrives without an
Authorizationheader is refused, from your own Mac too, and no setting changes that. - The token’s scope and its connection allowlist.
- The connection’s External Clients level: Blocked, Read Only (the default for a new connection), or Read & Write. The effective permission is the lower of that and the token’s scope.
- The connection’s AI Policy of Never, its Safe Mode level, and the Approval setting under Settings > MCP, which decides how often a client is asked before it reaches a connection.
claude.ai and app.cursor.com, so a page you have open there can call your local server from your own browser. The four gates above are what stands between such a call and your data, and the first one means a page without a token gets nothing.
Reads use the driver’s read-only mode where supported. Writes run only as approved, in the named database and schema. Statements that access files or run server-side code are refused. External query rules describe the limits.
Every call is written to the local activity log with the token behind it, statements as SHA-256 digests. See Tokens for the whole model.
Update checks
Sparkle fetcheshttps://updates.tablepro.app/appcast.xml once a day, and once for every check you start with Check Now or Check for Updates…. It is a plain GET with no query parameters: system profiling is off, so nothing about your Mac rides along. The user agent names the app and its version, and GitHub Pages, which hosts the feed, sees your IP address as it would for any web request. The Beta channel sends nothing extra: the one feed lists stable releases and betas, and your Mac picks between them. Versions up to 0.81 fetch https://raw.githubusercontent.com/TableProApp/TablePro/main/appcast.xml instead.
Updates download from the GitHub release named in the feed. Sparkle checks the feed’s EdDSA signature, and each build’s, against the public key inside the app, and refuses either one that does not match before it unpacks anything.
Automatically check for updates in Settings > General is on by default. With it off, only a check you start reaches the feed: Check Now in the same section, or TablePro > Check for Updates….
Download and install updates automatically is also on by default, so a new build downloads without being accepted first and installs when TablePro quits. The download is the same signed archive from the same GitHub release, and nothing is sent to fetch it. Turn it off to be asked each time.
The plugin catalog
At launch the app asks forhttps://updates.tablepro.app/plugins/v3.json, and its HTTP cache answers without a request while the last response is under ten minutes old. It asks the server again before it reports a plugin missing, and when it repairs a plugin that a newer app version rejected. The request adds no header of its own; GitHub Pages sees the IP address and the user agent URLSession sends. The response is cached on disk.
Installing a plugin downloads its ZIP over HTTPS from the URL in that catalog and checks it against the SHA-256 the catalog declares.
There is no setting that turns the catalog fetch off.
License checks and the team library
From Mac 0.86.0 (including 0.86.0-beta.1), Linux 0.1.0-beta.1 and the Windows preview, sign-in uses the account service athttps://api.tablepro.app. POST /v2/devices registers the machine ID, name, platform and app/OS versions. Account access tokens and signed plan files replace routine license-key checks. On Mac, the machine ID is the SHA-256 hardware identifier; on Linux it is HMAC-SHA256 of /etc/machine-id with app.tablepro.TablePro, with a stored random fallback; on Windows it is HMAC-SHA256 of the MachineGuid registry value with the same key. Refresh tokens use the Mac Keychain, the Linux keyring or Windows Credential Manager; Mac account tokens never sync.
Mac builds before 0.86.0 use https://api.tablepro.app/v1/license after a key is entered. Activation and seven-day revalidation send the license key, machine ID, Mac name and app/OS versions. Deactivation sends the key and machine ID. Device records and audit changes are covered by the account policy.
On Mac, publishing Team Library sends connection definitions and saved SQL to the service. Structured credentials are excluded; free text is not scrubbed. Current account builds use /v2/teams/{id}/library, pull changes at launch, every 15 minutes and after publishing, and publish changes rather than replacing the whole library. Older builds use /v1/license/library and pull at most weekly. Team Library describes access, and retention distinguishes normal team deletion from deleting its sole owner’s account.
Shared links
The Mac, Linux and Windows apps send nothing for sharing until you open a share sheet, which reads your limits fromGET /v2/share-limits on https://api.tablepro.app. Create Link sends POST /v2/shares with the title, note, database type and SQL, and for a result the columns the grid shows and its rows up to your plan’s limit. Masked columns are masked on your computer and sent with the name of their mask. Removed columns are not sent. The at-limit list and Revoke use GET /v2/shares and DELETE /v2/shares/{id}. Open in TablePro reads GET /v2/shared/{token} without the account, sending a password in X-Share-Password when the link has one.
A link’s page is served from https://s.tablepro.app. Its responses carry X-Robots-Tag: noindex, nofollow, noarchive, Cache-Control: no-store and Referrer-Policy: no-referrer. Views are counted per request, leaving out link previews and crawlers. A report sends its reason, details and optional email, checked with Cloudflare Turnstile, and the service keeps an HMAC of the reporter’s IP address, never the address. Shared Links lists what a link holds.
Crash reports
Mac 0.86.0 includes PostHog’s SDK with automatic crash capture disabled. Earlier Mac builds have no crash reporter. Linux sends no automatic crash reports. Local Mac logging uses OSLog. The iPhone and iPad app has no crash SDK; Apple’s separate Share With App Developers setting can supply crash reports and usage statistics through App Store Connect.Your database connections
The largest thing leaving your Mac is the traffic you asked for: the connection to your own server, plus whatever hops you configured, such as an SSH tunnel, a Cloudflare tunnel, a SOCKS proxy, or Cloud SQL Proxy. Cloud drivers reach their vendor’s endpoint by design, so BigQuery talks tobigquery.googleapis.com, Spanner to spanner.googleapis.com, DynamoDB to AWS, and Microsoft Entra authentication to login.microsoftonline.com.
Two helpers download on first use, and each is verified before it runs: the Cloud SQL Proxy binary from storage.googleapis.com against a pinned SHA-256, and the Copilot language server at a pinned version, against pinned SHA-256 digests of the archive and the binary.
What stays local
The execution log records every authorization decision the app makes, from the UI, an MCP client, the AI assistant, the import pipeline, or background maintenance. Each record holds a sequence number, a timestamp, the connection ID, the operation kind, which of those five callers asked, allowed or denied, whether it counted as a write, a SHA-256 digest of the statement, and the previous record’s hash. The statement text is never stored, and neither is a caller-supplied label such as an MCP client name or an AI session ID, only the channel it came through. The hash chain is tamper evident, not tamper proof: anyone who can write the file can recompute every hash after the record they changed. It is a JSON file in Application Support, no sync record type covers it, and no code path uploads it. It has no size cap. A file cut short mid-record, after a crash, is closed after its last whole record, with a copy of it as found kept asExecutionAudit.damaged-<date>.json. A file with no record TablePro can read is renamed to that name and kept, and a new chain starts beside it.
The table load history records how long each table load took, so a slowdown can be looked at after it happened rather than only while it is happening. One line per load, holding a timestamp, the app version and build, the macOS version, what started the load, the database type, whether the connection went through SSH, how long each phase took (schema, driver fetch, result conversion, grid reload, and the main run loop after it), how the load ended, any anomalies it reported, and four size ranges: rows, columns, result size, and open tabs. Sizes are ranges such as 101-1000 and 1-10MB, never exact counts.
Nothing identifying is in it: no query text, no rows, no database, schema or table names, no hostnames, ports or usernames, no connection names or IDs, no file paths, no error messages. It is a .jsonl file in Application Support, readable only by you, capped at 7 days and at 10,000 records or 5 MB, whichever comes first. No sync record type covers it and no code path uploads it. Delete the file to clear it.
Where it lives on disk
Every path here is under~/Library/Application Support/TablePro/ unless it says otherwise.
Passwords and secrets
Imported, synced and shared connections need local sign-in approval before using this Mac’s cloud credentials or Kerberos tickets. Environment variable references have a separate prompt. These approvals stay on the Mac that granted them. Secrets live in the macOS keychain, under thecom.TablePro service and in the data protection keychain: connection passwords, SSH passwords, key passphrases, SSL client key passphrases, TOTP secrets, plugin secure fields, AI provider keys, and older license keys. Account refresh tokens use a device-only Keychain item. Google sign-in refresh tokens for BigQuery and Spanner are kept in the same keychain under com.TablePro.GoogleOAuth, one per OAuth client ID. MCP tokens are the exception, stored as a salted SHA-256 hash; the plaintext is shown once at creation and never written down.
A connection can name a password source instead of using the keychain: a file, an environment variable, a shell command, 1Password, HashiCorp Vault, AWS Secrets Manager, or a Vault database role. Those resolve at connect time, on your machine, and are never synced, since a path or a command belongs to one Mac. A Vault database role sends your Vault token to the address it names, and the database password it gets back stays in memory.

