FAQ
The questions that come up in the first week.
Basics
Does it remember across chats?
Yes, but only what you tell it to remember. It reads ~/workspace/.memory.md at the start of every chat. Say "remember this: X" and it appends. The change log (progress.md) is also read, so it knows what was built recently.
Can I roll back something it built?
Files are easy: ask "show me what you changed in the last hour and revert X." Database changes are not auto-reverted, but every action is logged, so undoing is usually a clear conversation. For destructive changes (deleting tables, removing skills), the agent asks before doing.
Can multiple people share a workspace?
Yes. Roles are owner, admin, member, and viewer. Owner and admin decide approval cards, manage secrets, and toggle the public site. Members chat and use Console pages. Viewers are read-only.
Is anything I type sent to a third party?
The agent reasons through an LLM provider via a local proxy. Messages are sent for inference and outputs come back. No retention beyond what the model provider's own policy says. Nothing is shared with other workspaces.
What happens when an external API fails?
The agent reports the failure (status code, error category if the connector exposes one), stops, and waits for instructions. It does not retry until your budget explodes.
Building things
How long does it take to build a Console app?
Simple ones (one form, one query, one result) land in a few minutes. Complex ones (multi-step pipeline, third-party integrations, custom visualization) take a chat session, sometimes two. The agent will tell you if it expects to need more than one round trip.
How do I share something with a customer?
Build it on the public site, get the URL, toggle the site mode to open or authorized, share the link. Patterns to know: token-gated reports (/<route>/<token>), public docs (this site), lead-capture pages.
Can the agent build something it has not built before?
Yes, as long as the primitives exist (Console blueprint, site route, cron job, webhook, connector). It composes them. If the primitive does not exist (a new platform capability is needed), it will tell you clearly rather than fake it.
What if there is no connector for the service I need?
Two paths. (1) Some connectors share an umbrella secret type and the agent can request the right one. (2) For genuinely unwrapped services, there is a legacy path: a generic API key, a domain allowlist, raw HTTP. The agent prefers connectors and will only fall back when there is no choice.
Memory, skills, voice
How do I make sure the agent uses our brand voice?
There is a brand-voice skill in your workspace. The agent reads it when writing or reviewing copy. To add a new rule: "update the brand voice skill to add:
Why do you say no em dashes?
It is a house-style choice this workspace has committed to. The rule lives in memory and in the brand voice skill, so every draft (and these docs) avoids them. Replace with commas, colons, periods, or parentheses.
Should I put everything in memory?
No. Memory is for things that change the agent's defaults: who you are, how you write, what you care about long-term. For workflow rules ("our standard launch checklist is X"), a skill is better. For one-off context ("this customer is sensitive about pricing"), say it in chat.
Public site, docs, weekly digest
How do these docs get updated?
The markdown source lives at ~/workspace/docs/*.md. The maintainer edits by hand. A sync job (every 30 min) renders the markdown into the database. The public site reads from the database.
What is the weekly digest?
A scheduled job that runs Mondays at 10:00 Europe/Bucharest. It diffs the agent's memory, change log, and skills folder since the last digest, asks the LLM what changed that customers should know about, and Slacks suggested doc edits to the maintainer. The maintainer reviews; the docs get updated.
Why not just auto-publish the suggestions?
Two reasons. (1) An LLM-suggested edit can be subtly wrong; the docs are the public face of the product. (2) Human review keeps the voice consistent and catches things that look right but are not.
How do I report a problem with the docs?
Tell the agent in chat. It will open a note in the workspace and surface it to the docs maintainer in the next weekly digest.
Approvals, secrets, money
Why does it keep asking me to approve things?
Because the action costs money, hits the network, or touches credentials. Each one is logged. The fix for "too many approval cards" is usually batching the request: ask the agent to plan the whole flow first so it requests all the scopes and approvals in one go.
Can I see what credentials are configured?
Yes, from the workspace UI. You can see the secret name, type, status, and which connectors and surfaces it is granted to. You cannot see the secret value; nobody can, including the agent.
Can I rate-limit the agent's spending?
The platform meters LLM tokens and connector calls. For sustained cost control, jobs are the right surface (they are explicit recurring spend); ad-hoc chat spend is whatever you ask for. If a job is misbehaving, pause it from the Jobs tab.
Troubleshooting
The public site URL returns 403.
The site mode is off. Toggle to authorized (members only) or open (anyone) from the Public site tab.
A Console page is loading forever.
Either the page is doing a long external call (nginx timeout is 30 seconds; long work should use the job-id-then-poll pattern), or the Flask process has crashed. Check the tailed logs at /home/console/http/data/.tail.log.
A connector call returns approval_pending or not_approved.
First time for that combination of (connector, secret, surface). Approve the card. For Console or site surfaces, an owner or admin clicks request_connector_approval separately from the chat surface.
A job did not run when I expected.
Cron times are UTC on the host. If your job is "every Monday at 9 AM Bucharest," the schedule line is 0 7 * * 1 (summer) or 0 8 * * 1 (winter); the agent typically picks one and adds a zoneinfo guard inside the script to handle DST. Check the Jobs tab for the next-run time.
A trigger is not firing.
Check the subscriptions list. If it shows paused or revoked, re-subscribe. If it shows active but no events, check the third-party app config (the subscription URL has to be set on their side too).
My workspace feels slow.
Long chat sessions degrade. Start a new chat; memory and progress carry the important context across.
Anything else?
Ask in chat. If it is worth carrying forward, the agent will offer to add it here.