Bring knowledge in. Use it where you work.
Source connections bring selected information into Radcliffe. Assistant connections let an authorised tool ask Radcliffe for evidence. They solve different parts of the workflow.
Early-access requests are open. Product onboarding opens later.Two kinds of connection. One access boundary.
How it worksRead from your sources
Files, folders and authorised accounts
- Documents and knowledge
- Messages and work records
- Repositories and code
Use from your tools
Authorised assistants and coding clients
- Claude and ChatGPT
- Cursor and VS Code
- Other supported MCP clients
Connecting a source makes material available for processing. Connecting an assistant gives that client a way to request the evidence you may read.
Sources are where the knowledge comes from.
Choose the collection or account to read, authorise the relevant provider and check its scope. Source connectors read information; they do not edit the original documents.1
Files and company knowledge
Upload files or sync a laptop folder. Provider-specific setup also covers sources such as Google Drive, SharePoint and Notion.
Messages and work records
Connectors have their own scope and limits. For example, the Slack implementation reads permitted channel messages and threads, excluding direct messages.
Check readiness, then sync
The platform shows the connection’s state, last sync and problems that need attention. A catalogue entry may still need app credentials, consent or real-data approval.
Assistants are where you use the knowledge.
Radcliffe exposes an MCP server: a standard interface through which supported tools can search, fetch source passages, ask questions and request context.2
Conversations
Setup routes cover Claude and ChatGPT. Each person signs in to Radcliffe and authorises the client; managed accounts may also need administrator approval.
Coding tools
Setup routes cover Cursor, VS Code, Claude Code and Codex, among others. Install actions are offered where the client supports them; other clients need their own configuration.
Access stays personal
Searches and source reads use the authorising person’s current access. The connection does not grant company-wide access to the assistant.
A connection has a lifecycle.
Setup instructions explain how to begin. The connection state tells you whether authorisation and access have actually been established.3
Authorise
Complete the provider or client’s sign-in flow and approve the requested access. Copying an address is not authorisation.
Verify
Check that the expected source or client appears in the workspace. Test with a source you can already read before relying on the connection.
Manage or revoke
Source connections support management and disconnect flows. Authorised assistant clients appear as grants that can be removed.
Client setup guides
These are setup routes in the product, not a list of accounts connected to this website.
- Claude
- ChatGPT
- Claude Code
- Codex
- Cursor
- VS Code & Copilot
- Gemini CLI
- JetBrainsCompatibility check needed
- OpenClawCompatibility check needed
- Hermes
- Vercel AI SDKCompatibility check needed
- Other MCP clientsCompatibility check needed
Questions, answered.
Are all of these connections live on this website?
No. This launch collects early-access requests. The product contains connector and client setup paths, but a name on this page does not mean that a live account has been connected or validated for your company.
Does connecting ChatGPT or Claude import old conversations?
No. Authorising an assistant to query Radcliffe is separate from importing an exported conversation file. It does not automatically read your previous chats.
Is a model API key the same as an assistant connection?
No. A model API key configures a model used by the product. An assistant connection authorises an external client to call Radcliffe’s tools.
Why can a guide be available before a connection is ready?
A provider may require an app registration, administrator consent or a particular account plan. Hosted clients also need a reachable Radcliffe service. Your workspace’s setup state is the source of truth.
Product details
- Source implementations and permission models vary by provider. Connector code rejects content-changing requests. Credentials, company permissions and live-data gates are separate from a catalogue entry.
- The default MCP tool set is search, fetch, ask and context. Individual OAuth grants identify the caller. Client support, hosting and organisation settings determine whether setup can be completed.
- Connection management and OAuth grant revocation are implemented. Setup guides, installation actions and authorised connections are distinct states; no public page performs authorisation by itself.
What would your team ask first?
Tell us about the work and the information behind it.