seendiff Review (2026): Pricing, Features & Honest Verdict

seendiff is a free, open-source local diff viewer built for reading the enormous code changes AI agents produce, with separate tracking for what you have seen and what you have actually reviewed. Sharp idea, very early project. Best for: developers drowning in agent-written pull requests. Price: free, MIT licensed (free plan: yes). Rating: 6.5/10.

What is seendiff?

seendiff is a command-line tool that starts a local web server and shows you your branch’s diff in a browser. You run seendiff inside any git repository and it opens the full change set against origin/master, with every changed file shown whole rather than chopped into hunk fragments. It is a reader and nothing else: no editing, no conflict resolution, no comments.

The premise is stated on the homepage and it is the reason the tool exists. Coding agents produce diffs far larger than a human would write, and somebody still has to read them. It launched on Product Hunt on 29 August 2026 at rank 4, and the GitHub repository was created on 2 August 2026, so at the time of writing it is 27 days old. It is MIT licensed, written mostly in HTML with a Python CLI, and the commit history shows 39 commits from one main contributor, currents-abhishek, who lists Currents AI as their employer. The repository owner account, ACME.BOT, was created in July 2026 and has one public repository.

What Are seendiff’s Key Features?

Seen and reviewed are two different states

This is the whole idea and it is a good one. A block is marked seen automatically when it scrolls through your viewport. It is marked reviewed only when you press R on it, or f for the entire file. Every other diff tool collapses those into one concept, which is why you can scroll through 4,000 lines of agent output and end up with no idea which parts you actually read versus which parts your eyes slid past. A minimap runs down the side of the file so one glance tells you what is seen, what is unreviewed, and what you have not reached.

Review state survives a rebase

Close the window, come back tomorrow, and your progress is still there. Amend a commit or rebase the branch and it survives that too, which is the thing that usually destroys review progress in a web-based tool. State lives in ~/.local/state/seendiff and is pruned after 60 days, with a cache in ~/.cache/seendiff pruned after 30. Nothing else is written outside the repository, and --clear drops the state for the current branch when you want a fresh pass.

Fully keyboard driven, with a search that spans files

Arrow keys move between change blocks and files. n and N jump to the next and previous unreviewed block and cross file boundaries to do it. V marks every seen-but-unreviewed block in the file as reviewed, y copies a reference plus content, Y copies references for everything unreviewed in the file, s toggles side-by-side against unified, h hides fully reviewed files. Ctrl-F searches the file you are in and every other changed file at once. There are six themes with matching syntax palettes rather than one colour scheme bolted to every chrome.

A walkthrough skill so the agent explains its own diff

seendiff ships an agent skill alongside the CLI. Installed once, it lets a coding agent open a diff and write a step-by-step walkthrough of the code it just produced, anchored to the actual blocks. The framing on the site is the honest version of the AI coding argument: the agent is fast but lacks the context to make the design tradeoffs, so a trained pair of human eyes still has to be in the loop. This feature is the tool trying to make that loop cheaper rather than pretending it is unnecessary.

Built for diffs that are too big to open anywhere else

The rendering is virtualised, and the site claims a 40,000-line file scrolls smoothly. That number is the tell for who this is aimed at: nobody hand-writes a 40,000-line change. Flags cover the rest of the awkward cases. seendiff main takes any branch, tag or SHA as the base, seendiff HEAD~3 reads just the last three commits, --fetch is never implicit so the tool will not touch the network without being asked, and --section-min and --section-max control how the diff is chunked. --no-auto-seen turns off the scroll-marking if you would rather mark everything by hand.

How Much Does seendiff Cost?

seendiff is free. It is MIT licensed, there is no account, no paid tier and no telemetry, and there are no network calls outside localhost. The install is one command through uv, pipx, or plain pip into a virtual environment you keep. It needs Python 3.10 or newer and git, and it has four dependencies: FastAPI, uvicorn, unidiff and Pygments. Nothing to build, no config file, no daemon.

Compare that to what the code review category charges. Graphite is $20 per user per month on annual billing for its Starter plan and $40 for the tier above. CodeRabbit is $24 per user per month annually for Pro and $48 for Pro Plus, and the company has raised $143 million. Those tools do more, obviously: they run reviews, post comments, and integrate with your pull request flow. seendiff does not compete with them so much as it fills the gap they leave, which is the half hour where a human reads the change before it becomes a pull request at all.

There is one cost that is not money. seendiff is not on PyPI, so every documented install path points at a git URL. There are no tags and no releases on the repository either, which means there is no version to pin, no changelog to read, and uv tool install git+https://... gets you whatever was on the main branch when you ran it. For a tool that touches nothing but your local git state that is a manageable risk. For anything you would put in a team’s onboarding docs it is not.

Who is seendiff Best For?

Use seendiff if you run a coding agent hard and have started rubber-stamping its output because reading 3,000 lines in a GitHub diff view is miserable. The seen-versus-reviewed split is a real solution to a real problem, and thirty seconds of install cost is nothing against the risk of merging code nobody looked at.

Use it if you work offline or under rules that forbid sending source to a third party. Nothing leaves the machine, there is no account, and you can read all 39 commits yourself in an afternoon.

Skip seendiff if you need to share anything. There are no comments, no annotations and no way to hand your review to a colleague, so this is a solo reading tool by design. Skip it if you want a maintained dependency: 4 stars, no forks, one real contributor and a first commit 27 days ago is a project that could stop tomorrow. And skip it on Windows for now, because the project’s own PowerShell install block is labelled untested by its author. Credit for the honesty, but that is not a line you want to read while installing something.

Best seendiff Alternatives

Graphite. $20 per user per month billed annually for Starter, $40 for the tier above with unlimited AI. A full stacked-pull-request workflow with review tooling on top, built for teams that ship many small changes. Much more product, much more process, and a per-seat bill. Better if your problem is the review pipeline rather than the reading.

CodeRabbit. $24 per user per month annually for Pro, $48 for Pro Plus, $40 for the security product, plus $0.50 per agent minute for the Slack agent. It reads the diff so you do not have to and posts comments with one-click fixes. The opposite philosophy to seendiff: automate the review rather than help a human do it. Worth having both.

delta and difftastic. Both free, both open source, both terminal-native. delta gives you syntax-highlighted side-by-side diffs inside git, and difftastic compares syntax trees so it can tell a moved function from a rewritten one. Neither tracks what you have reviewed, which is exactly the thing seendiff adds, but both are mature, packaged properly and will still exist next year.

Final Verdict: Is seendiff Worth It?

The insight here is correct and slightly overdue. Agent-generated diffs broke the assumption that scrolling past code means reading it, and no mainstream review tool has caught up. Separating seen from reviewed, keeping that state through a rebase, and letting the agent walk you through its own work are three good answers to a problem that is getting worse every month. The keyboard map is thought through, the dependency list is four items long, and the whole thing runs on localhost with no account.

What you are not getting is a product. Four GitHub stars, no forks, no releases, no PyPI package and 39 commits from one person at a company that is not the one selling it. That is a promising weekend of engineering that has been iterated on for a month, not something to standardise a team on. Install it, use it for the next monstrous agent diff you have to sign off, and see whether the seen-versus-reviewed distinction changes how carefully you read. It costs nothing and it uninstalls cleanly. Just do not write it into anyone’s onboarding guide yet.