AI tools

The AI agent doesn't answer from a guess — it acts through a fixed set of capabilities. Each one either reads something, offers you a view to open, or stages a change for you to commit. This page is the trust reference: what the agent can do, and how each capability is gated.

How actions are gated

DynoTable is safe by default. Every capability falls into one of a few access classes:

• Read (silent): lightweight, read-only lookups, mostly local or table metadata. No prompt; they just run. A few read items from DynamoDB: on a table with no local index, the first schema or sample-value lookup samples up to 100 items with a Scan (read capacity), and a vector search is billed per byte searched.
• Read (gated): reads that hit AWS and cost capacity (a query, a single-item read). The agent asks before the first one, then remembers your choice per your permission mode.
• Propose: the agent emits a chip you click to open a tab. Nothing opens until you click, but preparing the chip can read: a table with no local index is sampled (up to 100 items) to pick its columns, and a Smart Table chip reads a sample of its join values to check the join resolves.
• Open (gated): opens or changes a tab immediately, behind the permission gate.
• Write (via staging): routes through staging; you review and commit. The agent never writes to DynamoDB directly.
• Export (gated): pulls results out to a file. An export to S3 asks every time, even in Full Auto.

Every gated action is approved according to your permission mode: Manual, Auto, or Full Auto, set per profile (see AI chat → Setting up AI). Every gated decision is recorded in an always-on, local audit log.

What it can do

Understands your tables and saved work

The agent reads your schema to answer questions: the tables under the active profile, their key schema, DynamoDB indexes, billing mode, your saved Smart Table and Workbench specs, declared relationships, and open tabs.

It also uses the local table index as context. All of this is read (silent), with no prompt. It reads locally, except when a table has no local index yet: the schema lookup then samples up to 100 items from DynamoDB to build one. To make that read your own call, index the table first from the Table dialog's Indexing section.

Reads your data

When it needs live data it runs read-only PartiQL or Workbench SQL (SELECT only, including JOIN, GROUP BY, and aggregates), reads a single item by its key, or counts the items matching a query. These are read (gated): they hit DynamoDB and cost capacity, so the agent asks before the first one. A vector search is the exception: it runs as read (silent), with no prompt, and DynamoDB bills it per byte the search examines and returns, plus one item read when it searches by an existing item. A visual filter query uses the same index and key condition you'd compose in the query builder.

Computes over a whole table, exactly

Ask for a count, a sum, an average, or a breakdown across an entire table and the agent computes it directly, with no need to export the data yourself first. The result is exact over the full table: it reads every matching item rather than sampling a page or two, so a "how many orders last month?" or "total revenue by region" answer reflects the real data, not a partial guess. It can do the same to produce a transformed export: reshape, filter, or aggregate a whole table into a new file in one step. This is gated: a full pass reads the table and costs read capacity, so the agent asks first.

Opens the views you ask for

The agent prefers to offer a view rather than take over your workspace. It emits a chip you click to open or focus a table (optionally filtered), the item editor for an existing row or a new one, a Smart Table, or a Workbench tab pre-seeded with SQL. That's propose, and nothing opens until you click. When you've clearly asked for it now, it can open or refine a tab immediately instead; that's open (gated).

Stages writes, never writes directly

There is exactly one way the agent changes data: it stages changes in the staging area for you to review and commit. That's write (via staging). It stages a create, update, or delete one item at a time, or a bulk edit or bulk delete of up to 10,000 items in one call.

When a turn ends, the agent shows one review chip for everything it staged in that turn (Review N staged changes), and Pending changes opens on the staged rows.

Exports and works with files

The agent can export an open tab or a one-shot SELECT to CSV, JSON, or NDJSON, and track an export in flight. It can also inspect, read, or run a SQL transform over a file you've already exported (or attached to the chat), which is useful when the data you want to work on is a local file rather than a live table. Exporting and transforming are gated; checking on an export or reading a file is silent.

It can also export a table tab to S3: the tab's visual query, to a bucket, prefix, point in time, and format it names. An S3 export runs and bills in AWS and can't be cancelled, so it always asks, on every call, even in Full Auto and even if you allowed exports before. The prompt shows the destination, the format, the point in time and what will be read, and offers only Deny or Allow for that one call; nothing is remembered. The agent can only write to a bucket in your own account. The MCP server refuses S3 exports.

Asks when a request is ambiguous

When it isn't sure what you mean, the agent asks a clarifying question with a small set of named choices and pauses the turn until you answer.