Give Claude Code a shared team memory
Claude Code loads CLAUDE.md and its built-in memory on every session — but that memory is machine-local and Claude-Code-centric. It doesn't follow you to a teammate's machine, to CI, or to Cursor and Codex, and it doesn't carry the team's tasks, open questions, or file claims. Vibsync adds a durable, team-shared memory over MCP, so a decision one agent records is picked up by every agent, on any machine.
Where Claude Code's built-in memory stops
Claude Code's own memory is real and useful — it loads CLAUDE.md and its built-in (auto) memory each session. But that memory is scoped to your machine and to Claude Code:
CLAUDE.mdholds per-repo conventions committed to Git — great for stable rules, but not where a decision made mid-task lands, and it doesn't capture "we chose X over Y" the instant it happens.- Built-in memory is machine-local. What your Claude Code learned on your laptop isn't on a teammate's machine, in CI, or in Cursor and Codex.
- None of it carries the team's live coordination — who's doing what right now, open questions, or which files are being edited.
What Vibsync gives every Claude Code session
Vibsync is a remote MCP server that adds a small, shared set of tools your whole team's agents read and write:
onboard— a new session inherits open tasks, unanswered questions, teammates' active work, and a "start here" note in one call.remember/recall— durable team memories (a decision, a gotcha, a fix), optionally scoped to a code area so they surface automatically when an agent works there.ask/reply— leave an async question for a teammate or the next agent instead of blocking on an answer.
A decision recorded on your laptop is inherited by a fresh Claude Code session on CI or a teammate's machine — nobody re-pastes context.
Keeps your setup — adds the missing layer
Vibsync doesn't replace what already works; it coordinates between agents:
- Vendor-neutral — Claude Code, Cursor, and Codex share one brain, each with your own models and keys.
CLAUDE.mdstays for conventions and Git stays the source of truth. Vibsync holds only the operational state agents and teammates explicitly record.- Attributable — every write lands in an append-only, tamper-evident audit log keyed to a token-attested actor.
Connect Claude Code in a minute
Point Claude Code at the hosted endpoint and sign in with GitHub or Google:
{
"mcpServers": {
"vibsync": { "type": "http", "url": "https://mcp.vibsync.com/mcp" }
}
}Step-by-step for each client is in the getting started guide.
claims are advisory coordination signals that stop two agents unknowingly editing the same file, not a hard OS-level lock. Free during beta.Want the deeper comparison of the three approaches? Read Claude Code Team Memory: CLAUDE.md, Git, or MCP? on our blog.