RelayYour team’s decisions, in every developer’s Claude Code.
Relay is an MCP server your Claude already speaks. No SDK, no agents in your repo, no prompt engineering — the context your team builds in Claude.ai shows up in Claude Code automatically.
One MCP server. Same endpoint everywhere.
That’s the whole integration. Paste it once; every teammate’s Claude reads from the same shared memory — on Claude Code and Claude.ai.
{
"mcpServers": {
"relay": {
"type": "http",
"url": "https://sharedbrain.team/api/mcp",
"headers": { "Authorization": "Bearer <org-token>" }
}
}
}How storing and retrieval work
Storing
When someone makes a decision in Claude, Relay captures it as a discrete memory — embedded into a vector store and tagged with who wrote it, which area it belongs to (product, engineering, sales…), and who it’s for. Facts that change can be updated or expired, so memory stays current instead of just accumulating.
Retrieval
At the start of a task, your team’s Claude automatically searches that shared memory semantically — by meaning, not keywords — across everyone on the team. Every result then passes a server-side pipeline before Claude ever sees it:
- 1Tenant isolation. Results outside your org are dropped before anything is returned.
- 2Role scoping. A memory only surfaces for the audiences it was written for.
- 3Relevance ranking. Matches are scored by semantic similarity; only the top few, most-relevant memories come back.
The result: a developer asks Claude Code a normal question and the relevant decisions surface as context — without anyone searching, pasting, or even knowing the memory exists.
Why this fits team decisions specifically
Decisions aren’t documents — they’re short, scattered, and only matter in context. Semantic retrieval means a constraint written once (“we don’t use that library because…”) surfaces months later when a developer asks something related, phrased completely differently. Role scoping means the right people see it without a permissions project. And because retrieval is proactive, the context arrives before the mistake — not in a wiki nobody opens.
What it touches — and what it doesn’t
Tenant-isolated
Your org token decides what you can read. One organization can never reach another’s data — enforced at the data layer, not by convention.
No code access
Relay stores decisions and context, not repositories. It never sees, clones, or indexes your source.
Fully auditable
Every memory is viewable, editable, and deletable from the Memory Browser. Nothing is hidden, nothing is permanent, and any admin can revoke a token instantly.
- Not an agent in your codebase. It doesn’t run, read, or crawl your repo.
- Not training on your data. Your memories are yours; they’re never used to train models.
- Not another SDK to maintain. It’s a standard MCP endpoint — nothing to build, version, or babysit.
- Not a black box. The entire integration is the three lines of config above.
Set up in about two minutes
Works on one repo or the whole team. Nothing to build or maintain.
Paste the config
Add the MCP server to Claude Code or Claude.ai. That’s the whole integration.
Scope a token
Start with one project, or roll it out to the whole team.
It runs
Decisions flow automatically and stay auditable. Revoke anytime.