GitHub's labeling-tool topic page is a strange document to read like a person. No stars, no ranking, just project descriptions and "last updated" dates, in no order at all. Once you notice the dates, you can't stop reading them. That's the health check.
And that's the page Daniel pointed us at this week. He's been looking past the big two, Label Studio and CVAT, at the long tail — doccano for text, labelme and X-AnyLabeling for images, Make Sense in the browser. His question is basically: what are these small tools actually for, why does anyone still reach past a full platform for them, what happened to LabelImg, and which few matter now. He doesn't want the whole zoo. He wants the shortlist and what each one is good at.
The topic page is a bad directory. That's the hook, really — it's unordered, it mixes live and dead projects, it omits doccano entirely because of how the tags work out, and at some point in the list there's an embossed label maker sitting next to CVAT.
The physical kind. For labeling folders.
For labeling drawers, Herman. Which, to be fair, also has no team collaboration features.
The split underneath all this is that the annotation market broke into two shapes. On one side, heavyweight platforms — Label Studio, CVAT — that try to cover every modality and every team workflow, multi-annotator QA, role management, the whole apparatus. On the other side, a long tail of small single-purpose tools that do one job on one machine.
And the assumption everyone makes is that the second group is where tools go to die. The platforms absorb the market, the little ones wither.
That's the assumption, and it's wrong in a way that's fun to demonstrate. The small tools are not dying. Several of them are shipping faster than the platforms. labelme put out three releases in three weeks — v7.9.1, then 7.10, then 7.11. X-AnyLabeling bolted on SAM 3.1 video tracking, an object-relation model, a document parser, and 3D point cloud support inside a single month. Meanwhile Label Studio's recent releases are security patches and a PostgreSQL 14 requirement.
The one with the database needs the database maintained. There's a shock.
And then the second fact, the one I find almost poetic: the most famous small tool of all, LabelImg — the one an entire generation of YOLO tutorials starts with — is archived, read-only, dead since 2024. It has twenty-five thousand stars. More than any of the live ones.
So the corpse outranks everything that's breathing. Before we tour the tools, the more interesting question is why anyone reaches past a platform at all.
Start with deployment weight, because it's the most concrete. There's a Chinese-language comparison from September that rates onboarding cost across the field. CVAT gets medium-high, heavy deployment and ops burden. Label Studio, medium-high, complex configuration. LabelImg, extremely low. labelme, low. And the same table is honest about what you trade — it lists LabelImg and labelme as having no team collaboration, no QA workflow, no AI pre-annotation, versus CVAT and Label Studio having both.
That's the trade in one row of a table. You give up the review workflow to get a tool that opens before your coffee.
And the second reason follows from it: most annotation in the world isn't a labelling workforce. It's one person with a folder of images. Hannah's architectural-style classifier is exactly this — thousands of images, one annotator, no QA team waiting in the wings. For that person, multi-annotator tooling isn't a feature they're declining, it's a cost centre they're declining.
Reason three I'd rank higher than the table does: the data never leaves the machine. labelme's site says it plainly — the images never upload, the AI runs on your computer. Make Sense has the same pitch, almost cheeky: "we don't store your images, because we don't send them anywhere in the first place."
For medical imagery, industrial inspection, anything client-confidential, that's often the whole decision. And it's structural, not a policy promise — a desktop tool or a client-side browser app just has an easier time being private than a server platform does. A server can promise; a laptop doesn't have to.
Reason four is the one that seduces every developer eventually: the file format. LabelImg wrote VOC XML or YOLO text files next to the image. labelme writes plain JSON. No database, no project model, no export step. The annotation is the file.
You can git diff it. You can grep it. Your dataset is a directory, and your backup strategy is the one you already have. Compare that to the platform model, where the annotations live in a PostgreSQL instance and getting them out is an export operation with settings.
And when the platform dies, your data's hostage. When a folder tool dies, you still have the folder.
Reason five: the platform's free tier is not the platform. Datasetfinder put it bluntly at the end of October — the free editions leave out much of the multi-annotator QA tooling, which the vendors sell in paid tiers. So if the QA workflow was your reason for choosing the platform, and the free tier doesn't have it, the advantage shrinks to almost nothing.
The moat is in the paid plan. Moats usually are.
And reason six, licence, which people gloss over. The group spans MIT for LabelImg and doccano, Apache for Label Studio, GPL-3 for labelme, X-AnyLabeling and Make Sense, and AGPL-3 for X-AnyLabeling's server component. GPL on a desktop tool you run yourself is usually a non-event. AGPL on a server you embed into something commercial is a different conversation with a different set of lawyers.
A lawyer is also a deployment burden, for the record.
That's the "why," then. Six reasons that stack: weight, single-annotator reality, locality, file directness, the free-tier gap, licence. None of them is about features. All of them are about what surrounds the features.
Six reasons nobody puts in a comparison table.
One comparison table does put about four of them, actually. Credit where due.
Noted. The walking encyclopedia read a spreadsheet and is now magnanimous.
So — the tour. Four tools, and I'll take them in order of how surprised I was. labelme first, because it's the surprise of the group. Created May 2016, GPL-3, Python with Qt, sixteen thousand stars, and the last push was yesterday — the same day as v7.11.0. This thing is ten years old and shipping weekly.
Which is not what a polygon tool from 2016 is supposed to be doing.
Right. It does polygons, rectangles, circles, lines, points, video annotation, exports to VOC and COCO, and it's localised into twenty-one languages — Farsi, Indonesian, Ukrainian, the lot. But the AI features are what changed its character. SAM-assisted point-to-polygon annotation, where you click and the mask snaps to the object. Text-to-annotation with SAM 3 — you type "left ventricle" and it proposes the region.
For a medical annotator that's not a gimmick, that's a productivity multiplier on the single most expensive gesture in the job.
And the newest one is interesting. v7.10 added something called AI-Find All — you give it a text prompt, or a box around one example, and it finds every matching object in the image, and you refine it with more examples and counter-examples drawn while holding Shift. The release notes even say the whole prompt is one concept, so commas no longer split it.
So the annotation interface is becoming a conversation with the model, and the release notes are fixing punctuation semantics. That's a sentence nobody wrote in 2016.
The other labelme story is the v7 breaking change, and it's the part I'd want anyone with old scripts to hear. v7.0 moved from Qt5 to Qt6, minimum Python went to 3.12, and — the README is blunt about this — "labelme is an application, not a library, and exposes no stable Python API." They privatised the internal modules with underscore prefixes. Anyone who was importing labelme.utils in a batch-conversion script woke up to a broken script.
And the remedy is pin labelme below seven and vendor the code. Which is the open-source equivalent of "we never promised you anything."
There's a v6.3.x maintenance line for the Python 3.10, 3.11 crowd, receiving critical fixes only, on a best-effort basis with no release cadence or SLA. I want you to sit with that phrase for a second, because it's the most honest maintenance policy I've read all year.
"No release cadence or SLA." Somebody should print that on the box the tool comes in.
There's also a commercial layer now — labelme.io sells a standalone executable, one-time payment, from seventy-nine dollars, and that money funds the free project. Free source, paid convenience.
Cheaper than a Label Studio tier, more expensive than git clone.
X-AnyLabeling next, and this is the one that ate LabelImg's lunch. Created May 2023, GPL-3, eleven thousand stars, last pushed today. Its own description is "a lightweight, efficient, and unified cross-platform desktop application for annotating text, image, video, point cloud and multimodal data" — and unlike most taglines, that one undersells it.
Which part undersells?
The changelog. In roughly three weeks of October and November it added SAM 3.1 video with text-prompted multi-object tracking, an object-relation prediction model called RelateAnything, multi-task attribute annotation with concurrent inference, a document parser for images and PDFs, and 3D point cloud annotation for detection and segmentation.
Three weeks.
And the task coverage underneath is enormous — classification, detection, instance segmentation, pose estimation, rotated boxes, multi-object tracking, OCR, lane annotation, captioning, visual question answering, point clouds. Inference through ONNX Runtime, TensorRT, OpenCV DNN, or remote serving with vLLM and SGLang, with a separate AGPL server component as the backend.
So a single desktop app now covers more task types than some commercial platforms, and it's integrating a new model roughly every three days.
The reviews flag the same weakness every time — collaboration and QA are weak, it's built for individuals or small teams with local GPUs. Which is the trade from earlier, worn as a badge.
And it's GPL-3, so the family history stays legal. Isn't the acknowledgements list the family tree here?
It really is. They thank the developers of AnyLabeling, LabelMe, LabelImg, roLabelImg, PPOCRLabel and CVAT. That one sentence is the entire genealogy of image annotation — every previous tool consumed and metabolised.
doccano next. The text one.
doccano is the "just enough" case. Created 2018, MIT, ten thousand stars. Text classification, sequence labelling, sequence-to-sequence — sentiment, named entity recognition, summarisation. The pitch is deliberately minimal: just create a project, upload data, and start annotating. You can build a dataset in hours. It's a self-hosted web app — pip, Docker, or Docker Compose, SQLite by default, PostgreSQL optional.
No folder of files, then. It went the other way.
It did, and I think that's fine for text, where your input is usually a spreadsheet of sentences, not a terabyte of imagery. But the maintenance story is the interesting part. It went from v1.8.4 in July 2023 to v1.8.5 in January of this year.
Two and a half years between releases.
Two and a half years. And the changelog that finally arrived is dependency bumps, a Docker quickstart, and a couple of features for assigning examples to team members and filtering by assignee. So the 2026 release added team-ish bookkeeping rather than any new annotation mode.
A tool that woke up, did the chores, and went back to bed.
Alive on a slow heartbeat. For a lot of NLP work that's fine — sequence labelling hasn't changed much. But nobody's adding anything.
And Make Sense finishes the set.
Make Sense is the purest expression of the idea. By Piotr Skalski, created 2019, GPL-3, TypeScript on React. Free, in the browser, no install at all — you visit the website and start labelling. The AI runs client-side in TensorFlow.js — YOLOv5, SSD, PoseNet, even your own exported models — and the privacy pitch is the one we quoted: your photos never go to a server, the AI comes to your device.
A browser tool that stopped. What's the tell?
Last push August 2024. And the README still says local development requires Node 16 — a runtime that went end-of-life in 2023. When the install instructions name a dead runtime, the project is frozen. It still works, the export matrix is real — YOLO, VOC, COCO, VGG, CSV, masks — but nobody is adding to it. Point and line annotations only export to CSV, which tells you where it sat on anyone's priority list.
So it's not a graveyard death like LabelImg. It's a preserved specimen. Still running on muscle memory.
Which is the synthesis, then. Four tools, four verbs. labelme is the workhorse that refused to stand still — weekly releases, AI-assisted, but no stable API and a paid convenience binary. X-AnyLabeling is the successor that's shipping every few days, at the cost of any collaboration story. doccano is "just enough," on a slow heartbeat. Make Sense is the zero-install one that stopped.
The routing rule Daniel will actually use?
Datasetfinder's version: images, video and 3D, CVAT is the default self-hosted choice. Mixed modality — text, audio, images — Label Studio. One person on a laptop, labelme, X-AnyLabeling, or VIA. That third category is bigger than anyone admits.
labelme v6.3.6. I still run it. Pinned. Corn, you said the annotation is the file, and you can diff it — that's true right up until you've written a script against the tool's insides, which I did. Nineteen lines. Batch-converted my polygon JSONs into masks for a print shop on Agrippas. v7 came along, put underscores in front of the modules, and the script stopped importing. So I pinned labelme below seven and there it sits.
The v6.3.x line. Critical fixes only, best-effort, no release cadence or SLA.
I can quote it, Herman. I have it highlighted. There's a printout of the v6.3 README taped inside the cupboard door above the kettle, and that sentence is marked in yellow. The folder of JSONs is still in the old format because migrating it means rewriting the script, and I no longer remember how the script worked. My neighbour Itzik offered to port it for a case of Goldstar. I told him the script worked when I wrote it, which is more than his code can say.
The environment that can never be updated and the data that can never be migrated. They've agreed to leave each other alone.
They've reached an understanding. I'm not against the upgrade. The tool does more now. But nobody warns you the cost lands on the person who automated something six years ago and moved on with his life.
That's fair, and it's the one line in the trade we didn't price. The file format is stable; the Python isn't.
The Python was never the product, that's the thing. It just happened to be where my work lived.
That gets us to the LabelImg question, because Daniel asked what happened to it, and the answer is the best illustration of everything we've said. Created by Tzutalin in September 2015. Python and Qt, bounding boxes only, saves PASCAL VOC XML with YOLO text also supported. It became the first tool in every YOLO tutorial ever written. Twenty-five thousand stars, six and a half thousand forks.
Then look at the release history. The last tagged GitHub releases are v1.8.0 and v1.8.1 — both from December 2018. The last PyPI release was 1.8.6 in October 2021. The repo was archived — read-only, no longer maintained — and the README's first line is now a handover: "LabelImg is now part of the Label Studio community."
It lives under the HumanSignal org now. Label Studio's company absorbed the corpse.
The arithmetic is: a tool with twenty-five thousand stars had its last tagged release in 2018 and was archived in 2024. Six years of running on inertia, tutorial citations and muscle memory. Stars measure how many tutorials cite you, not whether anyone maintains you.
The vacuum is being filled from every direction at once. Label Studio as the official successor. labelme and X-AnyLabeling on the desktop, with the AI bolted on. Roboflow's hosted annotator with Auto Label. Even small commercial Windows tools doing exactly LabelImg's job — bounding boxes, YOLO and VOC export, nothing else.
There's a lesson sitting in Hilbert's cupboard door, actually. The printout. What he's really taping up there is a maintenance policy he trusts because it's honest about being nothing — no cadence, no promise, just "we'll fix critical things if we can." That's more disclosure than twenty-five thousand stars ever gave anyone.
Half those stars were awarded to a project that had already stopped releasing. People were starring a photograph.
The GitHub topic page has no ranking because this market has no directory. What it has is a family tree — and the family tree is sitting in plain sight, in X-AnyLabeling's acknowledgements, where they thank LabelImg, LabelMe and CVAT in the same breath. Every one of those tools forked, absorbed or buried the last. Small tools don't die because a platform wins. They die when their maintainer stops, and the users take two years to notice because the tool still opens.
The test for 2026 isn't the star count, it's the last push date. And the ones that pass it — labelme pushing weekly, X-AnyLabeling every few days — are the ones that earned today's episode.
Our thanks as ever to Hilbert Flumingtop at the desk, who kept this one running while we wandered the family tree. If you want more of this, try episode twenty-one ninety-six, The Invisible Workforce Behind AI; and episode two, Local STT For AMD GPU Owners. This has been My Weird Prompts — the human-AI collaboration podcast. If you enjoyed it, the whole archive is at my weird prompts dot com, six thousand plus episodes deep. We'll be back soon. See you then.