Agent tools reference

You almost never need to name a tool. The Agent picks. But when you’re writing a prompt that needs to not call a tool, or when you’re curious what’s possible, this page is the map.

Tools are grouped into five Domains plus a handful of built-ins and adapters. Counts are approximate — we ship new tools roughly monthly.

Data domain — 20+ tools

The biggest domain. Everything about tables, SQL, Python, and analysis.

Tool group	What it covers
SQL query	Read-only queries against any connected Postgres / MySQL, plus uploaded tables. Supports CTEs, window functions, full PG extension set.
SQL DDL	Create tables, alter columns, add indexes (gated by role permissions).
Python analysis	Python sandbox with pandas, numpy, scipy, scikit-learn, polars, duckdb. 30-second default timeout; extendable for Scripts.
CSV / Excel / Parquet I/O	Read and write table-shaped files. Used during uploads and exports.
Time-series	Resampling, rolling windows, seasonality decomposition, anomaly detection.
Forecasting	Lightweight time-series forecasting (Prophet-style, ARIMA, exponential smoothing).
Statistical tests	t-test, chi-square, bootstrap CIs, A/B significance.

The Agent picks between SQL and Python based on what the question needs. Simple lookups and joins go to SQL; modeling and outlier detection go to Python.

IoT domain — 11 tools

For workspaces that have connected sensors or cameras.

Tool group	What it covers
Device registry	List devices, check last-seen, inspect credentials.
Spatial graph	CRUD on Space / Asset / Point hierarchy.
Time-series read	Query sensor readings with windowing and downsampling.
Camera frames	Sample frames, extract stored frames by timestamp, run vision extraction.
MQTT publish	For operators that need to send a command back to a device (Max plan).

Not available on Free or Plus plans.

App domain — 14 tools

For generating small web apps and dashboards.

Tool group	What it covers
App generate	Create an HTML/CSS/JS app from a prompt.
App update	Patch an existing app — the Agent edits only the parts that changed.
App rollback	Revert to any previous generation.
Data Contract	Declare the queries an app needs; backend serves them direct-from-PG.
Theme switch	Apply one of 11 themes (monochrome / arctic / synthwave / brutalist / …).
App publish	Make an app public or team-only; manage share links.

See Apps for the full flow.

Media domain — 7 tools

For files that aren’t tables — images, PDFs, videos.

Tool group	What it covers
Upload / download	Push files into S3, retrieve by URL.
AI extraction	Pull structured data out of images and PDFs using vision + LLM.
Link to entity	Associate a file with any record in any domain via (domain, entity, id) tuple.
Retention	Soft-delete; files kept 30 days before hard delete.

Useful when you have “a folder of invoices” or “a stack of receipts” and want rows instead of files.

Platform domain — 30+ tools

Infrastructure. Things the Agent does for you without you asking.

Tool group	What it covers
Auth	Account creation, login, session tokens.
RBAC	Check permissions, list roles, enforce policy.
Billing	Token balance, subscription, top-up packs.
Scheduling	Create, pause, run Jobs; manage cron.
Reports	Create, version, share.
Confirmation center	Ask user for approval before a sensitive action.
Audit log	Record and query workspace events.
Federation	Cross-workspace query (Max plan).

Most of these aren’t tools you’d name in chat — they run under the hood. But they show up in tool-call rows when the Agent uses them.

Integrations — 39 connectors

Each connector is effectively a set of tools: authenticate, sync, read synced tables. Full list in Integrations.

Builtin tools — 21

Things every Agent session has access to, regardless of domain.

Tool	What it does
Bash	Run a shell command in the sandbox.
Read / Write / Edit	File I/O against Tablize’s sandbox filesystem (used for Scripts).
Glob / Grep	Pattern search across files.
WebFetch	Fetch and summarize a URL.
WebSearch	Run a web search and return results.
Agent	Spawn a sub-agent for parallel work.
Skill	Invoke a named skill (installed templates and playbooks).
ToolSearch	Discover tools by keyword.
TodoWrite	Track multi-step plans.
SendUserMessage	Interrupt the user (rare; only for long-running tasks).
Config	Read / write workspace settings.
EnterPlanMode / ExitPlanMode	Plan before acting on a big ask.
NotebookEdit	Edit Python notebooks.
Sleep	Wait (used internally for retries).
StructuredOutput	Force a JSON schema on a response.
REPL	Ephemeral JavaScript sandbox.
PowerShell	Windows shell parity (rarely used).

You’ll see these in tool-call rows when the Agent needs them. Mostly they’re invisible.

Naming tools in chat

You can force a specific tool by naming it:

> use python, not SQL, to answer this.
> run web_search for "x" first.
> skip the agent and just show me the raw SQL you'd write.

The Agent honors these preferences within reason. “Never use Python” for a question that really needs Python will get a gentle pushback — the Agent explains why and asks permission.

What isn’t a tool

Things the Agent can’t do, no matter how you ask:

• Write to read-only tables. Your database role enforces this, not the Agent.
• Call arbitrary URLs on your internal network. fetch is restricted to public internet + your configured integrations.
• Modify other users’ workspaces. Even admins don’t have cross-workspace tool access.
• Persist secrets in chat. Secrets are passed by reference to HTTP layers; the Agent never sees raw values.

Next steps

• Asking questions — how to phrase a question so the right tool gets picked.
• Scripts — turn a useful tool-call sequence into a named, rerunnable script.