🚧Known limits in the beta

About this page#

The beta is a working product, built and checked part by part. Some parts aren't finished, and some have only been tried with stand-ins that follow the vendors' documentation, not with the real services. This page lists the limits you may meet; it's updated as they're fixed.

Not yet tried with the real services#

Each of these works end to end with stand-ins. The first real checks happen with the first beta users: if one of them doesn't work for you, write to us (below).

• Live runs on the Claude API, with your organization's own key (BYOK).
• Signing in with Google or Microsoft.
• Connecting tools by signing in to them (OAuth), and renewing those sign-ins automatically.
• Reading a connected tool's actions from its MCP server, and testing them.
• Fetching a Reference's web page when a run starts.

Partly done#

• A Reference reaches the agents as inline text, or as a web page where link fetching is switched on for your organization. A PDF or a file on your computer is given as an address only.
• A Skill runs only when it's linked to a tool through a Connector with an MCP server; local (stdio) servers can't run on the platform.
• Invitations aren't sent by e-mail: an admin copies the link and sends it.
• Every organization has the same three roles; areas and per-project members come later.
• Deleting your account or your organization, or getting a copy of its data, isn't self-service: your admin writes to support.
• The Process X-ray reads .txt, .md and .docx files; not PDFs yet.
• On a phone, the Canvas shows a notice instead of the editor: use a tablet or a computer to edit.

Not built yet#

• Using files and tools on your own computer.
• Company structure: areas, company rules and access requests.
• Insights, history and suggested improvements across your runs.
• Sharing workflows with other organizations, and partner packs.
• Exporting to environments other than Claude Code.

By design#

• A run's answers aren't kept on our servers: they're shown to you and dropped 15 minutes after the run ends. Copy what you need to keep.
• Simulation is free and never needs a key: use it to check a workflow before a Live run.

When something goes wrong#

When a run fails, it says what failed, at which step and how to fix it, on the Canvas and on the run's page under Runs.