Skip to main content
Scopes are your specialist AI configurations — your “digital workforce.” Instead of one agent that tries to do everything, you create focused experts, each scoped to a domain with just the instructions, context, and tools it needs. A scope is reusable: tag it on Home or drop it into a workflow step.
The Scopes list showing a published scope with search, view toggle, and visibility filters

​
The Scopes list

The list page manages your whole roster:
  • Create Scope — start a new specialist, or Duplicate an existing one to reuse its setup.
  • Tree View / List View — switch between the supervision hierarchy and a flat, searchable table.
  • Search and visibility filters (All / Shared / Private) to find scopes fast.
  • Each row shows the scope’s description, its tools & integrations, and its access. Badges flag a Draft (never published) or unpublished changes.
  • The row menu lets you Publish / Deploy Changes, Duplicate, or Delete.
When you create a scope you give it a name and optional purpose, choose a visibility (Shared with groups in your org, or Private to you and people you invite), and optionally assign a parent scope for supervision.

​
Configuring a scope

Open a scope to configure it. Everything is organized into tabs, and a live test conversation panel on the right lets you try changes as you make them.
A scope's configuration page: Instructions tab with a description, ordered instructions, and a live test panel
A short description of the scope’s role, an ordered, drag-to-reorder list of instructions (its operating rules), and an optional expected output format. This is the scope’s system prompt.
Select the integrations this scope may query, down to specific databases and schemas, plus tool-actions for SQL, BI, coding, and messaging connections. Keeping this tight is what makes a scope focused.
Choose the slice of your context available to this scope — driven by its connections, plus any specific notes, files, saved queries, or links you attach. A coverage view shows what’s included.
Enable the general tools this scope may use (for example web search, file publishing, and connected app actions). Connection-scoped SQL access lives in the Connections tab.
Let the scope save and reuse memories from conversations, and search or manage what it has stored.
Lists every conversation with the scope — searchable and filterable by time and user.
Browses conversation files and files the scope has published.
Define test cases — a task plus an expected output judged by an LLM — and run them to track whether the scope still behaves as intended after changes.
Exposes the published scope to external AI tools like Claude, Cursor, and ChatGPT — see Connecting AI tools. Copy per-host install snippets, toggle Core Tools (chat, get_context), and choose which integration tools appear in the MCP catalog.
Controls ownership, Shared/Private visibility, group-based Use vs. Manage permissions, external invites, and safety guardrails.
Click Versions in the left nav to open the version history side panel. Every publish creates a version — review a diff of what changed between versions and restore a previous configuration if needed.

​
Publishing and versioning

While you edit, changes stay in a draft. Click Publish Scope to make it live; after that, edits show as unpublished changes and you Deploy Changes (with a review of what changed) to roll them out. The Versions tab keeps the full history so you can compare or roll back.

​
Supervision and hierarchy

A scope can have a parent that coordinates it, letting a supervisor delegate to focused sub-scopes. Use Tree View to see and manage this hierarchy — a good pattern when one broad request is best answered by several narrow experts.

​
Using scopes

Once published, a scope is available on Home — open the Scopes dropdown and tag it to bring its expertise and context into the conversation. Scopes are also the building blocks of workflows, where each AI step runs a scope.
Give each scope a narrow focus and only the context it needs. Smaller, focused scopes are faster, cheaper, and more accurate than one broad one — and easier to evaluate and trust.