Qlaude lets Claude Code work with a model server you run yourself
A small open-source launcher points Anthropic's coding agent at your own model server, so you can try it without an Anthropic account.
The short version
- Qlaude is a single shell script that wires Claude Code to your own server
- No Anthropic subscription or account is needed; Windows is not supported
- vLLM has answered Anthropic's Messages API itself since v0.11.1
- A hosted provider with a compatible endpoint can stand in for your own server
- Setup is five questions, then five self-checks you can rerun with qlaude doctor
Why it matters
If you want an AI coding agent but do not want your code going to an outside service, this shows a route that avoids it. The catch is that results depend on the model you serve. A smaller local model may handle agent work less well than a hosted one.
What Qlaude does
Claude Code is Anthropic's coding agent that lives in the terminal. It reads your code, edits files and runs commands. Normally it talks to Anthropic's own service. Qlaude, a project on GitHub, redirects it to a model server of your choosing.
As github.com reported, the project's pitch is that you need no Anthropic subscription or account. You also do not have to own a GPU (a graphics chip used to run AI models) or run a server yourself. A hosted provider that offers an Anthropic-compatible endpoint can take the place of your own server.
It is a launcher, not a server. The repository does not run a model for you. You bring three things: a model server that is already running, its web address, and an access token.
Why the connection works
The reference server in the project is vLLM, an open-source program that serves language models. Since v0.11.1, vLLM answers Anthropic's Messages API directly. That is the request format Claude Code uses to talk to a model.
Because of that, Claude Code can connect straight to vLLM with no translation layer in between. The project says anything that speaks the same route should work. It lists a test of Claude Code 2.1.325 with vLLM 0.33, plus a second stack using llama-swap and llama.cpp serving a Gemma model.
The listed tests were run in November 2026 on Linux and macOS. Windows is not supported.
Setup in five questions
Qlaude is a single script written in POSIX sh, a very portable shell language. You install it with one pasted command. The installer checks each downloaded file against a published checksum. The project is candid that this protects against corruption, not against a compromised host, since the checksums arrive over the same connection.
A short interview then asks for:
- the endpoint URL of your server
- your access token, hidden as you type
- the model, picked from the ones the server lists
- the context size, meaning how many tokens the model can see at once
- the reasoning effort, meaning how hard the model thinks before it answers
The settings go into a config file readable only by your user. The token is never printed or passed as a command-line argument. You set no environment variables by hand, because the launcher exports what Claude Code needs each time it starts.
Checks and daily use
At the end of setup, five checks run. They confirm that the needed tools exist, that the server returns a reply, that Claude Code runs, and that a tool call makes a full round trip. Tool calls matter because that is how an agent edits files and runs commands. You can repeat the checks any time with qlaude doctor.
After that, typing qlaude opens Claude Code against your endpoint. qlaude config shows your saved settings without revealing the token. You can also pass arguments straight through to Claude Code after a double dash.
The project also lists a fix for a misleading error. A message saying the model may not exist is, according to the docs, almost always caused by a trailing /v1 on the address. For vLLM itself, the docs point to flags that make tools work, including a tool-call parser and a maximum model length.
Who should try it, and what to expect
This is for self-hosters and anyone with code they would rather keep off outside services. Our hosts flagged it as very early: it had one point and no comments on Hacker News at last look, and no price is stated.
The main trade-off is quality. Claude Code is built around Anthropic's own models, and the agent's results will follow whichever model you point it at. A small model running at home may struggle with multi-step tasks that a hosted one handles well.
The software you are redirecting also has its own terms, so read them before you rely on this setup for work. Removal is simple: delete the script, delete the config folder, and on macOS remove the stored keychain entry. The project is licensed under Apache License 2.0.
What to watch
- How well agent-style coding holds up on smaller local models compared with hosted ones
- Whether the project gains users and testing beyond the Linux and macOS stacks it lists
- Whether Windows support is ever added
Try it
Install it on a throwaway project, run qlaude doctor, and compare the agent's output with your usual setup before trusting it with real code.
Sources and links
- Reportinggithub.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by github.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
