The compose file publishes Postgres with a known password
chat web app for teams, sass with user management and ratelimit, support chatgpt(openai & azure), claude, gemini and ollama model
At a glance
- What is it?
- swuecho/chat is a Go and browser chat application for teams with user management, a rate limit and conversation snapshots, built as one binary that embeds its own web assets. Its deployment defaults are written for a trusted network, and its documentation links point at somebody's laptop.
- Who is it for?
- Treat the defaults as a local development setup, because that is what the comments say they are. Two things to change before this faces a network: the database port published with a password written in the file, and the first-registered-user-is-administrator rule, which decides who controls the model keys.
- Can I use it commercially?
- Yes. MIT 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?
- Yes. The repository last received commits 42 days ago.
- What is it written in?
- Mainly Go, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 19, 2026, and from our analysis. They are not legal advice.
Star history
- swuecho/chat573 ★ · -2
| Repository | 2026-09-15 | 2026-11-19 |
|---|---|---|
| swuecho/chat | 575 | 573 |
-2 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The mobile app link points at the author's own filesystem
The README has two halves, one in Chinese and one in English, and both link the mobile app the same way. The link target is an absolute path on a Mac, with a user's short name, a development folder and the repository name in it.
/Users/hwu/dev/chat/mobile/README.mdThat link cannot resolve for anyone else, and it is the only pointer to the Flutter client that the documentation gives. The mobile directory does exist in the repository, so the fix is a relative path, but as published it is a broken link that also discloses the author's local directory structure.
The rest of the documentation is properly linked: guides for adding a model, for using a local Ollama server, for the OpenAI-compatible gateway, and for deployment, each in both languages, plus a community forum and a separate document explaining the difference between snapshots and live chatbots.
That last distinction is the one worth reading before you deploy, because the same application does both.
The compose file publishes Postgres with a known password
The deployment file runs two services, the application and a Postgres 14 database, and it is explicit about what it does.
- PG_PASS=thisisapasswordPOSTGRES_PASSWORD: "thisisapassword"The database runs as the postgres superuser, and its port is published to the host rather than kept inside the compose network, next to a health check that runs the database's own readiness probe every five seconds. The application publishes its own port too.
The API keys are placeholders, and the comments next to them say not to change them if you do not have keys for that provider, which is a reasonable way to keep the file copyable. The database password has no such warning, and it is the one value in this file that will still be there after you deploy it somewhere public.
The one other good habit here is that the application waits for the database to report itself healthy before starting.
The shipped rate limit is the one the file tells you not to use
The rate limit is a single environment variable, and both the README and the compose file give it a default and a warning.
The default is one hundred model calls per ten minutes. The comment beside it says to set it to zero if the server is on a public network, and to only raise it in the admin panel for trusted users.
So the file ships a value its own comment describes as wrong for production, which is defensible for a development default and easy to miss when you copy the file.
Two other knobs are visible but commented out. One caps streaming per word for the first two hundred words of a response before switching to per line, which is a detail about how output is chunked rather than about cost. The other points the OpenAI client at a proxy. Neither is documented elsewhere in the README, so the only place to learn they exist is the file itself.
One binary, two toolchains, one architecture
The image build runs in three stages and the second one is the interesting part.
A Node stage compiles the browser front end, then a Go stage downloads the module dependencies, copies the front end's output into the Go project and compiles the binary with the C library disabled. The front end is therefore not served by a separate process or a CDN. It is embedded in the same executable.
The final stage adds an unprivileged user, copies the binary and the timezone database out of the Go image, and points an environment variable at that copy, because the minimal runtime image has no timezone data of its own. A health check hits a path under the API on port 8080.
One line in the build is a constraint rather than a choice. The architecture and operating system are both written into the compile command, with the architecture fixed to 64-bit x86. So the image you build is for one architecture, and on an ARM host it runs only through emulation.
The first person to register is the administrator
The rules are short and one of them is a security decision rather than a feature.
The first user who registers becomes the administrator. Everything about model keys, the rate limit per user, and the choice of a separate model for generating conversation titles lives behind that account.
So on a deployment that is reachable by anyone, the order of registration decides who controls the application's spending and configuration. The rate limit guidance in the deployment file is the same instinct from the other direction: treat the network as hostile unless you have made it otherwise.
Elsewhere the same rules show the application's priorities. A conversation starts with a system message that acts as the prompt, the four most recent messages are sent as context, and any conversation can be published as a static page or resumed later. Snapshots live in a directory with full-text search, in English.
Two languages, two lists, and a description that promises more
The README presents the same feature list twice, once in Chinese and once in English, and the two lists are not identical.
The English list describes the built-in OpenAI-compatible gateway in two bullets, one for revocable virtual keys used to call configured models and one for the proxy behaviour with per-key limits, auditing and time-limited samples of requests and responses. The Chinese list compresses the same two ideas into a single line and a follow-up line.
The repository description promises more than either list. It names Azure alongside OpenAI, and Gemini alongside Claude and a local model, while the English rules mention only OpenAI, Claude and Ollama, and the deployment file carries keys for OpenAI and Claude only.
A separate guide for adding a model exists in both languages, so the supported set is a configuration question rather than a fixed list, which is probably why the summary drifted.
Deployments track the latest tag unless you pin them
The compose file pulls a published image rather than building one, and the tag it uses is the moving one.
image: ghcr.io/swuecho/chat:latestThe same line carries two alternatives in comments: an image under a different registry namespace, and a note to use a version tag instead for better stability, with an example of an early version number.
So the documented default is that restarting a deployment months later gives you whatever was published most recently, which for a project whose releases run every few months is a real difference in behaviour. The release history is a sequence of three-part version numbers, so pinning is possible, and the comment shows the author knew it.
A platform configuration file for one specific hosting provider sits at the repository root next to the compose file, which suggests the container route is the primary one.
Editorial conclusion
Treat the defaults as a local development setup, because that is what the comments say they are. Two things to change before this faces a network: the database port published with a password written in the file, and the first-registered-user-is-administrator rule, which decides who controls the model keys. Also fix the documentation link that points at a path on the author's own machine.
Frequently asked questions
What is swuecho/chat?
A Go web chat application for teams, with user management and a rate limit, supporting OpenAI, Claude and locally hosted Ollama models. The browser front end was copied from an existing open-source chat web app, and the API layer was written from a node version of a pull request on that same project.
How much conversation context does swuecho/chat send?
By default the latest four messages are included in context, and the first message of a conversation is treated as the system prompt. Both behaviours are stated in the rules section of the README.
Who becomes the administrator in swuecho/chat?
The first user who registers. The admin panel is where model keys, per-user rate limits and the model used to generate conversation titles are configured, so registration order decides who holds those controls.
How is swuecho/chat deployed?
With the bundled compose file, which starts the application and a Postgres database, or from a published container image, which the file pulls using the latest tag by default with a comment suggesting a pinned version for stability.
Can swuecho/chat keep a record of past conversations?
Yes. Conversations are kept as snapshots in a directory with full-text search in English, and any conversation can either be published as a shareable static page or continued afterwards.
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/swuecho-chat)Star history chart for your README
[](https://hysenlabs.com/projects/swuecho-chat)Was this useful?