Frequently asked questions

Getting started

Do I need to know SQL? No. You upload or connect data, approve the metrics that were proposed from it, and ask questions in plain language. Metric definitions are editable if you want that level of control, but nothing requires it.

How much data do I need to try this? One spreadsheet. A single CSV with a date column, a category column and a number is enough to create a working agent and see the whole flow.

Can I start without connecting a live system? Yes — upload a file. You can move to a connector later without rebuilding the agent, as long as the columns still line up.

Data

Which file types are supported? CSV, TSV and Excel workbooks. For a workbook with several sheets, choose the sheet to use during import; the selected sheet becomes one dataset.

What happens to columns beyond the AI profiling cap? They are still ingested and still queryable. They simply arrive without an AI-generated name, description or proposed metric — you can write those yourself.

How do I refresh data from a connected source? Three ways: on a schedule (hourly, daily, weekly), on demand when a question needs it, or manually with Refresh now.

What if the source's columns change? The refresh detects the drift and reports it instead of loading a table your metrics can no longer read correctly.

Can I undo a cleaning step? Yes. Cleaning transforms are reversible.

Are my credentials safe? Connection secrets are stored encrypted and are never displayed back to you after saving.

Agents

Why did my agent refuse a question? Because it fell outside the scope you defined, or it needed a metric the agent is not allowed to read. Widen the catalog selection or add the metric to its scope — deliberately, not reflexively.

Why is my agent stuck in Draft? Draft becomes Tested automatically after a successful run. If it will not go Live, open the Readiness tab: scope, metric documentation or refusal review may still be incomplete. Evaluations are quality guidance, not a separate hard publishing gate.

My agent has tools connected but still says it has no data. Connected HTTP or mail tools are not a data source. Going Live requires an approved workspace metric, an enabled per-user upload slot, or a live spreadsheet connector. Tool availability, approval mode and connection scope also depend on workspace tool settings and the agent's bindings.

The agent gave a wrong number. What is the right fix? Look at How I worked this out to find which metric it used, then fix the metric definition or its description. Correcting the wording in chat fixes one conversation; fixing the metric fixes every future one.

Can I make several agents share the same data? Yes. Metrics and catalogs are workspace-level; several agents can be scoped to the same catalog without duplicating anything.

Publishing and clients

Do my clients need an account? They must sign in to enter the Portal. They can accept an email invitation or sign in and redeem a shareable join link to join as consume-only workspace members. An owner or collaborator can also issue a targeted link for a published agent; a matched client need not join the whole workspace. Public result snapshots have separate links and normally do not require sign-in.

Will my clients see changes as I make them? No. Clients run a frozen snapshot of the agent's client-facing configuration; your edits reach them only when you republish. Owners and collaborators preview the live workspace version. Data is always live. Tool bindings and approval modes are fixed in the snapshot, while tool specifications, connection scope and credentials are checked at run time.

What can a client not do? Create or edit agents, view datasets, metrics, skills or connections, reach any builder page, or see agents you have not published.

Why is Publish disabled? The agent is not Tested or Published, or its required readiness items are incomplete. Draft, Paused and Deprecated agents cannot publish. Scope, metric documentation and refusal review are required; evaluations are quality guidance. Publishing itself is available on every plan, Free included. Publishing the Portal also requires a link name and available account capacity, and must be done by the workspace owner or a collaborator.

How many portals can I publish? There is one Portal per workspace. How many workspaces and Portals your account can create and keep online depends on its plan capacity. Adding an agent to an existing Portal does not consume a separate allowance. See Plans and credits and the in-app usage page.

What happens to published agents if I downgrade? Downgrading does not delete existing resources. Existing Portals generally continue serving existing clients, though the account may temporarily exceed the new plan's limits; creating resources or starting refreshes may be blocked. You can take a Portal or agent offline for free.

Is sharing a result the same as Portal access? No. Portal access requires sign-in and either workspace membership or a valid targeted agent share; the agent must still be published. A result share is a public, read-only snapshot of one conversation and can have its own access code, expiration and revocation. Revoking or expiring that link also stops viewing and downloading it; it does not grant Portal access.

Can a public result link accept follow-up questions? Not by default. The person creating the share can explicitly enable follow-ups, which use a temporary identity per visitor and platform-configured visitor limits, with an optional limit for that link. Follow-ups query current data and use credits from the agent's account, not the visitor's account. Visitors can only read: they cannot write data, export files or perform external actions. Follow-ups can be disabled after sharing but cannot be enabled later for that link.

Billing

How are credits calculated? From actual model usage — input and output tokens at the model's rate — plus applicable units for generated media such as video or spoken-audio duration, document pages or slides and images. It is not one credit per question, and rates can change as platform administrators update provider pricing. Current rates and allowances are listed on the pricing page and in-app usage page.

Why does a call show a reservation first? Before a metered call starts, estimated usage is reserved from the account credit pool. At the end, the call settles against actual usage: unused reservation is returned, failure releases it or settles at zero, and usage above the reservation is added. In-progress reservations and successful settled usage appear separately on the usage page.

Do multiple workspaces have separate credit pools? No. Workspaces, members, Portal clients and approved follow-up questions under one billing account share an account credit pool. Plan credits apply to the current daily or monthly window; purchased top-up credits are a non-expiring account balance. Workspaces cannot split these into separate allowances.

Am I charged for failed requests? No. Failed calls do not count toward final successful usage; their reservations are released or settle at zero.

Do unused credits roll over? Plan credits refresh at the applicable daily or monthly window. Purchased top-up credits do not expire. Owners and collaborators can also set a monthly spending limit for an individual member.

Still stuck?

Report a wrong answer directly from the message — it reaches the builder with the question and the answer attached. For anything else, use the in-app feedback option.