Claude Code proposes every Git operation over MCP and nothing runs until you approve it in FluxGit. You see the exact diff, a risk level and a restore point first, and one click undoes anything you approved.
Three steps. Every command below comes straight from the README of the open-source MCP server, so what you paste is what actually ships.
Download the desktop app for macOS, Windows or Linux. This is where the approval cards appear. If you only want the free read-only shell without the app, one command puts fluxgit-mcp-sidecar on your PATH:
cargo install --git https://github.com/fluxgit-hq/fluxgit-mcp-server fluxgit-mcp-sidecarClaude Code is an MCP host, so the generic block works as-is. Paste it into your MCP config; the FluxGit app pre-fills the absolute sidecar path and gateway address for you under Settings, Agents / MCP, Quick Connect:
{
"mcpServers": {
"fluxgit": {
"command": "/absolute/path/to/fluxgit-mcp-sidecar",
"env": {
"FLUXGIT_GATEWAY_ADDR": "127.0.0.1:14660",
"FLUXGIT_MCP_AUDIT_LOG": "/optional/path/to/audit.jsonl"
}
}
}
}Say "commit what you changed" in Claude Code. Claude calls repo.brief to orient itself, then sends operation.preview.commit with its reason and the exact files it wants to stage. A card appears in FluxGit; you read it and click Approve or Reject. That is the whole loop.
No more discovering a reset after it already happened. Every Git write Claude wants arrives as a card with a plain-language reason, the exact diff, a risk level and a restore point captured before anything runs.
Approved a batch and regret it? The completed card offers one-click undo through the guarded reset flow, and every approval keeps its restore point in the Safety Timeline.
Orienting in a repo through raw git output cost 1,886 tokens on a 44-commit test repo; one repo.brief call answered the same questions in 607. Measured, not estimated. That is context Claude keeps for your actual task.
Public repositories are free. A $39 founder license, one-time, adds private, self-hosted and local-only repos. The MCP shell itself is open source under Apache-2.0.
No. FluxGit speaks the open Model Context Protocol, so any MCP host connects the same way: Codex, Cursor, or whatever you switch to next. Nothing in the setup above is Claude-specific.
No. The MCP server exposes no write capability at all; every mutation is a proposal, and the only path to a Git write is your click in the desktop app. That is construction, not policy.
That is the point. The card explains what the operation does in plain terms, flags the risk, and shows the way back before you approve. You learn Git by reviewing real decisions instead of memorizing commands.
I built FluxGit so the machine asks first. If you want the full tool surface and the wire contract, the MCP feature page has all of it.