Let your AI agents use Git. Without fear.
The agent never gets write access to your repository. It proposes a change; you approve the exact diff in FluxGit, with its risk level and a restore point captured first, before anything runs. One click undoes an approved batch. Underneath sits a full safety-first Git cockpit: visual history, Trinity 3-way conflicts, semantic diff with honest fallback, and Fleet Radar for many repos.
The real FluxGit interface, on synthetic demo data.
These are real screenshots of the FluxGit desktop app, captured with synthetic demo repositories, branches and commits only — no real customer or local repository data.
The whole idea, in 90 seconds.
The agent proposes, you approve, your own risky clicks stay guarded, and a full Git cockpit sits underneath — one silent walkthrough of the real app.
Two ways to make Git safe for agents.
There are two bets on how version control should handle AI agents. One is to let the agent commit freely and contain the fallout: sort its edits into branches, review afterward, roll back when it overreaches. GitButler does that well, and it is a reasonable bet if you trust the agent to write.
FluxGit takes the other bet. The agent never holds write access at all. It proposes a change; FluxGit shows the exact diff, a risk level and a restore point; nothing touches your repository until you approve it in the app. Approve before it runs, instead of reviewing after it already did.
Contain-and-review-after trusts the agent to write and gives you a good cleanup. Approve-before-it-runs never hands the agent write power in the first place. Every mutation is a proposal with a preview, a risk level and a restore point, and the only path to a Git write is your click in the desktop app.
If you would rather the machine ask first, that is the whole product.
Built for the moments Git gets dangerous.
Most Git tools make commit and push easier. FluxGit focuses on the expensive moments: conflicts, rebases, resets, dirty worktrees, lost commits, monorepos, submodules and multi-repo drift.
Founder license for early users: one-time payment, lifetime updates. Checkout opens at launch — request beta access to register interest.
View founder licenseOne cockpit for the Git work that normally breaks flow.
FluxGit combines a visual Git graph with safety rails, recovery, conflict resolution, semantic diff fallback, optional AI-assisted commit workflows, MCP for agents with human-approved writes, and local-first privacy controls. Using Claude Code, Cursor or another MCP agent? The 60-second setup.
MCP write-with-UI-handshake
Your AI agent proposes. FluxGit's approval modal shows the diff, risk, restore point. You decide. Eleven operations covering the full agent cycle: branch, commit and push included, plus merge, rebase, discard, reset, patch, worktrees and multi-step plans approved as one unit. And it saves the agent's context: one repo.brief call replaces the 10-command orientation sweep, 607 tokens vs 1,886, measured.
Semantic diffs
Tree-sitter structural diffs for 60+ languages. Moved code stops reading as churn, agent proposal cards open with the structural summary first, and unparseable files fall back, labeled, to the plain patch.
Safety rails
Risk levels, explicit decisions, safer alternatives and contextual recovery before destructive Git operations.
Flux restore points
Recovery-oriented checkpoints for dangerous history operations, designed to make rebase and reset less scary.
Reflog recovery
Turn local HEAD movements into visible rescue actions, including recovery branch creation after reset or branch deletion.
Trinity conflicts
Guided 3-way conflict workspace for base, current, incoming and final result review.
Predictive conflicts
Beta preflight branch integration checks to warn about likely merge conflicts before the user starts a risky operation.
Commit Studio
A larger focused commit surface for staging, reviewing diffs and composing commits without fighting cramped side panels.
The full toolbox
Blame, per-file history you can step through commit by commit, bisect, merge strategies, interactive rebase, content search across the repo and any file's contents at any commit. The basics are all here, wrapped in the same safety rails.
Fleet Radar
See which selected local repositories are dirty, behind, ahead, divergent or need attention without opening them one by one.
AI Cortex
Provider-aware AI-assisted commit workflows when configured, with local-first secret handling and normal Git still usable without AI.
Monorepo tools
Partial clone, sparse checkout, focused branch filtering and performance-conscious repository refresh policies.
Parallel workspaces
Run several agents at the same task in parallel worktrees, compare the attempts side by side, merge the winner into your checkout and clean up the rest. Submodules and parent-child repos ride the same guarded flows.
Safety Timeline
Restore points and HEAD movement merged into one recoverable history, so risky operations always have a way back. Agents can read it too, read-only.
Designed around real workflows, not isolated commands.
Branch status, remote drift, dirty submodules, reflog hints and Fleet Radar attention signals.
Checkout, discard, rebase, merge, cherry-pick and reset flows explain consequences before continuing.
Reflog recovery, restore points, conflict panels and contextual recovery avoid silent destructive guesses.