Architect conversations idle for more than an hour stop taking new messages. The conversation stays readable. A button starts a new conversation, and the new one carries over the objective. The Architect re-reads the workspace brief, so the new conversation continues from where the last one stopped.

Imports list every sheet in a workbook, including sheets that could not be read. Unread sheets appear greyed out, with the reason. The sheet list is no longer cut off at 50 sheets. Import emails name their own sheet and list any sheets that were not read.

Fixed: a workbook that had two sheet names differing only in punctuation (for example “Q1 2024” and “Q1-2024”) could fail to import.

The Architect now reports circular references between actions. Alerts about repeated re-runs no longer claim a loop they did not detect, and they are sent once per table instead of once per row.

The invite page offers Google and Microsoft sign-in. If you are signed in with the wrong account, the page says so and offers Sign out.

Fixed: the signup and invite pages showed the stray word “hidden”.

Fixed: the import queue list did not scroll.

The managed-actions panel on the table page shows each action’s inputs and outputs according to the action type.

Billing: an action run is now charged only when it did something: it changed data, sent an email or web request, wrote rows, or queued an import. Failed runs and runs that changed nothing are no longer charged. A network action is charged once per run, not twice.

Billing: the Architect chat now keeps its prompt cache for one hour instead of five minutes. A cache write costs 2× the input rate instead of 1.25×, but a conversation writes its cache much less often. Chat turns that use web search are now billed for the whole turn. Agent actions keep the 1.25× rate. The pricing page shows the current rates.

The Architect now works inside a workspace. Each workspace has an Architect tab. The tab holds the workspace brief: goal, status, open questions, decisions and reference files. Each conversation belongs to one workspace, and the conversation list shows only that workspace.

The workspace home page is now an Overview. It shows the goal, the current status and pinned tables. To pin a table, use its Manage menu.

Conversations are tasks. Each one has an objective, can be marked done, and is grouped by workspace in the list. The chat panel opens on where the build stands, not on the full log.

You can download reference files, and text files preview in place.

Load action:

• It processes every sheet of a workbook. It no longer stops at 12 sheets.

• It reads up to 2,000,000 cells per sheet.

• It skips an oversized sheet and names it, instead of refusing the whole file.

• It reads macro-enabled workbooks (.xlsm).

• Testing a load action runs a real trial import. The trial stops for review in the table’s import queue and writes no rows.

A workbook that cannot be read now tells you why, instead of showing a generic error.

Live lookups now re-run whenever any action changes the data they read. A re-run that changes nothing no longer triggers further re-runs.

Files attached to Architect conversations are sent more efficiently, which lowers the cost of long conversations with attachments.

Line spacing across the app is more comfortable to read.

The get tool no longer adds the parent context block on the first read of a path. Pass context: true when you want it.

New Architect features:

• A brief for each workspace: goal, status, decisions and questions.

• Conversations with an objective, which can be marked complete.

• Undo for the last turn. It proposes deleting what that turn created, and you approve it.

• Calculations over workspace files.

• Reading spreadsheet formulas.

• Saving research notes as reference files.

Tables can carry a description and can be pinned for the whole workspace.

New API reads: the full structure of a workspace in one request, and a public pattern library at /pattern.

Low-credit warning emails now link to your billing page.

Creating many rows at once is faster when live lookups are involved.

Maintenance release; no change you will notice.

The action page is redesigned to match the table page. It shows the action’s models on both sides, the destination table and the connected tables. One click moves between the action page and the table page.

Actions on the canvases are shown as tiles with the action type’s icon.

A network action’s write strategy is now set in its own dialog.

Import queue: the table toolbar has an Imports button that shows how many imports wait for review. You review imports there. It replaces the per-file transform pages.

The workspace page no longer has a Models tab. The team inbox is now the Email tab.

Grid number formatting: years have no thousands separator, decimals are never rounded, and currency shows two decimal places.

The Load action now appears in the app.

Fixed three ways a canvas could save a change you did not intend: a single click on a connection dot, a drag from a read-only dot, and removing a match column without a confirmation. Removing a match column now asks first.

Stronger security headers on the app. Security updates to dependencies.

New Load action. It imports a spreadsheet into a table through that table’s import queue. It replaces the automatic-import switch on child tables, which is retired.

From a table, you can create a parent table or a child table, with the link between them, in one step. You can also create an action that is already attached to a table, and list the tables connected to an action.

Public price list: /pricing shows the current storage, action and AI prices.

The usage page shows AI spend that is not tied to a table as its own line, so the per-table amounts add up to the monthly total.

For the Generate document action, the choice of Excel or Word is now a setup step.

A model now belongs to exactly one action. The standalone workspace model endpoints are removed.

The Architect can take more steps in one turn.

Fixed: two actions with the same name could not both be created.

Fixed: on an empty table, a reference column could not be made required or unique.

Security improvements to sign-in sessions and to API response headers. Existing sessions keep working until they expire.

Security fixes to password reset and email verification.

Password-reset links now expire after one hour and work only once. Links issued before this release no longer work.

The table page has a new Table view. It shows parent tables on the left and child tables on the right; click a table to move to it. The page switches between Data and Setup, and Setup switches between Table and Action.

The Action view also shows the actions that use the table: lookups, networks, joins, generate actions and agents.

The setup canvases scroll like a document.

New API endpoints list a table’s parent and child tables, create a parent table and its link in one request, and list the actions that use a table.

The Matrix action type is renamed Join.

Security updates to dependencies (web app) and to the MCP server’s runtime.

A table’s own address opens its setup canvas, with columns and connections on one page. Connections are shown as a stack of cards, with each card’s position in the run order.

You browse actions in a window laid out like the public library. It shows each action’s inputs and outputs.

Fixed: moving between connections on the canvas could clear a field’s link to a column.

Fixed: filtering the action list by scope failed.

Fixed: the usage page could count a charge twice when a charge was retried.

Fixed: the low-balance warning could count a run twice.

The column that a Join action matches on cannot be deleted while the join uses it.

Security updates to dependencies (API, web app and MCP server).

Actions are protected against runaway re-runs. When one change sets off an unusually repetitive chain of re-runs, the chain is stopped and reported. Large legitimate operations run to completion.

Renaming a column now updates the connections, conditions, match filters and references that use it.

Deleting a column or a table is refused while an action or an export template depends on it, and the message says what depends on it.

Billing:

• A low balance no longer stops runs that are already under way.

• The usage page reports exactly what was charged.

• Agent actions reuse cached history between steps, and cached input is billed at its own rate.

Waiting actions resume more reliably.

Every file field in the API accepts an uploaded file reference. Agents and scripts can use the upload flow for any file, for example a Generate document template.

Imports: when a sheet’s header row is ambiguous, the import now asks instead of guessing.

A table’s import settings show the header names and value rewrites that the table has learned from past imports, and you can delete them.

Action reliability fixes: conditions, network actions that are replayed (they no longer create duplicate rows), and waiting actions.

You can add, edit, delete and reorder columns on the connection canvas.

Fixed: percentage cells showed values 100 times too small. For example, 88.89% showed as 0.8889%.

Fixed: in a cell’s menu, the Download button could be pushed out of view.

Double-clicking a node on a canvas opens it without zooming.

New Generate document action (formerly called Export). It fills an Excel or Word template from table rows. Excel templates keep their headers, merged cells, formulas and macros, and you map table columns to Excel columns.

Agent actions can read attached Word documents, spreadsheets, CSV, plain text, JSON and SVG files.

An agent action’s limit is now a maximum cost per run, in credits, enforced during the run. Ordinary runs are no longer refused on a healthy balance.

Lookups:

• You can set a sort order, so “first” means first by that column.

• A lookup with a blank match value now fails instead of matching more rows than intended, and match inputs are now required.

Action test forms now enforce required fields.

One column can feed two input fields.

The last admin of a workspace cannot be removed or demoted. The option is shown disabled, with the reason.

Closing an account now revokes its API keys.

New in the web app:

• Matrix (now Join) and Generate document actions.

• A 30-day daily report for each action.

• A workspace error log.

• Publishing a table to the public library.

• A workspace settings dialog.

• One Structure page that replaces the separate Columns and Connections pages. Old links redirect.

Fixed: the Rename dialogs for tables and workspaces also showed publishing and access settings, so a table’s visibility or a workspace’s access could be changed by accident.

Fixed: controls you are not allowed to use are now shown disabled with the reason, instead of disappearing or failing after a click.

A connected mailbox now shows when it was last resynced and what the resync recovered. It also shows work that is stuck.

Fixed: after a failed save, a file cell could serve the previous file under the new file’s name.

You can read an automated import’s decisions for each file.

In Architect conversation history, approval cards now appear where they were raised.

API: maintenance release; no change you will notice.

Architect conversation history shows the real outcome of each past action, and no longer shows a success tick when the outcome is unknown. It opens on the newest messages, with a way to load earlier ones, and shows an error if the history fails to load.

Storage billing now charges per day as intended, and new storage rates apply from this release.

Balance notifications:

• Accounts that have made a deposit get a warning email before their balance runs out.

• If work stops because the balance is empty, you get an email, and another when the work resumes.

• Your account shows whether automatic top-up is set up, failing, or not enabled although a card is saved.

Fixed (v2.21.1): the low-balance warning was not sent.

Fixed: re-running selected rows in bulk failed silently while it reported success. If part of a bulk re-run fails, the dialog now shows which rows ran and retries only the rest.

Platform emails are now sent through a new email provider. Their content is unchanged.

Fixed: images in emails were broken, and some emails were sent twice.

A connected mailbox that paused for a low balance resumes as soon as a deposit settles. Before, this could take up to an hour.

Error messages from AI providers no longer repeat your data in action logs.

Fixed: the handling of cell values when many rows are created at once.

Security updates to dependencies.