Unbody archived itself and left a six-service retrieval stack behind
The Supabase of AI era. A modular, open-source backend for building AI-native software — designed for knowledge, not static data.
At a glance
- What is it?
- The AI-native backend's README declares the repository archived and forwards readers to Adapt, while the repository itself still holds a complete MongoDB, Temporal and Weaviate reference stack with two configuration mistakes worth reading before you run it.
- Who is it for?
- Judge this repository on what it still is: a complete reference stack for a retrieval backend, frozen on 2026-04-14 and forwarded to Adapt by its own authors. A team that needs a maintained AI-native backend should start with Adapt rather than this.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Activity is slowing. The repository last received commits 7 months ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- unbody-io/unbody525 ★ · +2
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| unbody-io/unbody | 523 | 525 |
2 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The README declares an archive that GitHub's own flag does not set
The file opens with the word archived in its title and a banner stating that the repository is archived and no longer maintained. Everything after that is a redirect. The company says the work now happens under Unbody Labs and names Adapt as the direct continuation: a provider-agnostic AI memory and learning framework, described as lightweight and under 200KB, able to run in the browser or on the server. It gets links to its own repository, its documentation and a live demo.
The original setup instructions still exist, folded inside a collapsed block labelled Original setup instructions. A reader arriving at the file has to expand a disclosure element before seeing a single command. Meanwhile the archived status is self-declared. GitHub's own archive flag is not set on this repository, the last push was on 2026-04-14, and no GitHub releases have ever been published. The banner is a statement of intent rather than a lock on the door, which matters if you were planning to file an issue and wait.
package.json calls the package private and unlicensed next to a LICENSE file
The manifest names the package unbody, pins it at version 0.0.1, sets private to true and license to UNLICENSED, and leaves both description and author as empty strings. The repository around it is published under Apache-2.0 and carries a LICENSE file among its top-level entries.
These are two different subjects and both can be true at once: a private application package that never went to a registry, sitting inside a permissively licensed source tree. Nothing in the manifest reconciles them, and nothing in the file explains the split. For anyone planning to vendor the source, the LICENSE file is the authority. For anyone planning to depend on a published package, the manifest says there was never one. The empty description field matters more than it looks, because npm and most registries render it directly in search results.
Six Compose services start before a single document is indexed
The development stack is defined in one docker-compose.yml with six services. MongoDB runs the 4.4.6 image started as a single-node replica set named rs0, with an init script mounted from the repository. Redis has no configuration at all beyond always restarting. Temporal runs the admin-tools image with ports 7233 for the server, 8233 for the UI and 60896 for metrics. Weaviate is built from a Dockerfile inside the repository rather than pulled. An img2vec-neural service and a python file-server complete the set.
Two commands bring the whole thing up and then start the Nest application on top of it:
docker compose up -d
yarn startThe store is pinned to an old 4.4 series release while the ORM dependency is mongoose ^8.10.0, so the database is considerably older than the client library written to talk to it. That gap is workable but it is the first thing to check when a replica set behaves strangely.
Weaviate listens on 0.0.0.0 with anonymous access switched on
The vector database service sets AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED=true and starts with --host 0.0.0.0 --port 8080 --scheme http, publishing 8080 and 50051 directly to the host. QUERY_DEFAULTS_LIMIT is set to 25. No credential appears anywhere in the service block, and nothing restricts the listener to loopback.
This is a laptop stack, and the file is candid about why it is built that way: an extra_hosts entry maps host.docker.internal to host-gateway so the vector database can call back into services running on the host. The trade is that a store holding indexed document chunks answers unauthenticated queries from anything able to route to port 8080, including other containers attached to the same unbody network. Copying this Compose file onto a server without changing that line is the failure mode to plan around.
The Weaviate module list is written twice and the two values disagree
The same service block assigns ENABLE_MODULES twice. The first assignment names a single module, text2vec-huggingface. Four lines later the identical key is assigned the full list: text2vec-huggingface,generative-unbody,img2vec-custom,reranker-custom,multi2vec-custom.
Only one value can survive, and a YAML mapping keeps the last assignment it reads. The earlier line is dead configuration, a leftover from when the service loaded one module. It is harmless when the later line parses as expected and actively confusing when you are trying to answer a simple question: is the generative module enabled? The file answers twice and gives two different answers. Note also that three of the five modules, generative-unbody, reranker-custom and multi2vec-custom, are custom modules built from a Dockerfile in this repository, so they exist nowhere else.
The environment file is called .env in one file and .env.local in the other
The header comment of .env.example tells you to copy the template to .env. The setup instructions in the README tell you to run cp .env.example .env.local and edit that instead. Two names for the same file, given by two files in the same repository, with nothing to say which one the application actually loads.
The template itself is small and reveals the shape of the deployment. It carries OPENAI_API_KEY, two local storage root directories for public and private files, matching public and private base URLs, SERVER_BASE_URL pointing at host.docker.internal:3000, and IMG2VEC_BASE_URL pointing at localhost:3456. The mix of host.docker.internal and localhost inside a single template is deliberate rather than sloppy: the application runs on the host while the image and vector services run in containers, so each setting names whichever side it has to reach.
Indexing progress is only visible through a dashboard you refresh by hand
Adding a data source is one command, which package.json maps to ts-node src/cli:
yarn unbody-cli source addEverything after that step happens somewhere the repository cannot show you. Indexing runs as Temporal workflows, and progress is observed in the Temporal UI at http://localhost:8233/, where the README is explicit that the dashboard needs a manual refresh to display the latest state. There is no polling endpoint described, no log file named, and no completion marker.
Once indexing finishes, the instruction is to follow the README in a separate examples repository. So the last mile of the workflow, deciding what a finished index actually contains and how to query it, lives outside this project entirely. Budget for that when you evaluate it: the retrieval backend is here, and the worked example of using it is not.
Editorial conclusion
Judge this repository on what it still is: a complete reference stack for a retrieval backend, frozen on 2026-04-14 and forwarded to Adapt by its own authors. A team that needs a maintained AI-native backend should start with Adapt rather than this. A team that wants to read working Mongo, Temporal and Weaviate wiring for document retrieval will get value from the Compose file alone. Before running anything here on a shared machine, remember two things: Weaviate answers unauthenticated queries on 0.0.0.0, and the manifest marks the package private and unlicensed.
Frequently asked questions
Is unbody-io/unbody still maintained?
No. The README carries an archive banner and states that the repository is no longer maintained, and the last push was on 2026-04-14 with no GitHub releases published. GitHub's own archive flag is not set, but the authors state the position plainly.
What did Unbody's authors point the project to as its successor?
Adapt, a provider-agnostic AI memory and learning framework that they describe as lightweight and under 200KB, running in the browser or on the server, with its own repository and documentation at adapt.unbody.io/docs.
What does it take to run Unbody locally?
Node.js LTS 20 or 22, Docker with Docker Compose, yarn rather than npm, and an OpenAI API key. The stack also expects a Temporal dashboard reachable on port 8233 once the containers are up.
How do you add a source and watch it get indexed?
Run yarn unbody-cli source add, then open the Temporal UI at http://localhost:8233/. The dashboard has to be refreshed manually to show the current state, and no completion marker or log file is named.
What license is Unbody released under?
The repository is published under Apache-2.0 and carries a LICENSE file at the top level, while package.json sets private to true and license to UNLICENSED at version 0.0.1. The two describe the source tree and the package separately.
Does the Unbody Compose stack need a GPU for image embeddings?
No. The img2vec-neural service sets ENABLE_CUDA to 0, so image vectors are produced on CPU. It maps host port 3456 to the container's 8080, and the image embeddings path is enabled through the custom img2vec-custom Weaviate module.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/unbody-io-unbody)Star history chart for your README
[](https://hysenlabs.com/projects/unbody-io-unbody)Was this useful?