The MCP server
MCP (Model Context Protocol) is the standard way Claude connects to outside tools. Relay exposes one MCP endpoint that every teammate’s Claude points at.
The four tools
The endpoint gives Claude exactly four tools. Two are invoked during normal use:
team_memory_searchReadSemantically searches the team’s shared memory and returns the few most-relevant items, already filtered to what the caller is allowed to see. This is the call recall runs.
team_memory_addWriteSaves one or more new memories, tagged with author, area, and audience. Supports a bulk mode. This is the call a save runs — only ever on an explicit, intentional save.
team_memory_updateWriteReplaces the content of an existing memory when a fact has changed — so memory stays current instead of piling up.
team_memory_deleteWriteRemoves a memory. Nothing is permanent; any memory can also be deleted from the Memory Browser.
What runs when
The tools don’t fire on a schedule — they’re invoked by Claude in response to what you do. In short: search runs on retrieval (automatically on Claude Code, skill-driven on Claude.ai), and add runs only on an intentional save — never per message. The Retrieval and Saving pages cover the exact timing.
Connecting & authentication
Each teammate connects with an org-scoped token that identifies their organization and role. How that connection is established differs by surface:
Both carry the same verified identity (tenant + handle + role); the difference is only how the credential is presented.
{
"mcpServers": {
"relay": {
"type": "http",
"url": "https://sharedbrain.team/api/mcp",
"headers": { "Authorization": "Bearer <org-token>" }
}
}
}