Install and run CoderAI

Quick start

The repository quick start builds all backends, activates the virtual environment, then starts the server on http://127.0.0.1:8776.

git clone git@git.nexlab.net:nexlab/coderai.git
cd coderai
./build.sh all
source venv_all/bin/activate
python coderai

First login: open http://127.0.0.1:8776/admin, log in with admin / admin, then change the credentials and create real API tokens.

The published packages

Everything CoderAI ships is a public container image on GitHub Container Registry — nothing to build. :latest moves only after an image has completed a real request on real hardware; every version is tagged too (:0.2.21).

Package	What it is
`ghcr.io/nextime/coderai`	The full CoderAI: every capability and engine, the Web Studio and admin UI, cluster head/node, llama.cpp with the RPC backend and `rpc-server`. One layer, ~28 GB.
`ghcr.io/nextime/coderai-images` · `-video` · `-text` · `-embeddings` · `-ocr` · `-tts` · `-stt` · `-voice` · `-audio` · `-faceswap`	One capability each, on the full core (9–10 GB).
`ghcr.io/nextime/coderai-speaker` · `-tts-xtts` · `-stt-nemo` · `-stt-crisper` · `-ocr-paddle`	Stacks with their own venv (pyannote, Coqui XTTS, NeMo, CrisperWhisper, PaddleOCR), 7–8 GB.
`ghcr.io/nextime/coderai-llama` · `-vllm` · `-engines` · `-engines-kt`	llama.cpp (with RPC), vLLM (with Ray worker mode), the native MoE engines and ktransformers as CoderAI pods.

Linux

docker pull, one script

docker pull ghcr.io/nextime/coderai:latest
curl -fsSLO https://raw.githubusercontent.com/nextime/coderai/master/packaging/linux/run_oci.sh
chmod +x run_oci.sh
./run_oci.sh --nvidia -d      # or --vulkan / --all; then http://localhost:8776/admin

The script installs itself as coderai-docker; coderai-docker --upgrade pulls the production branch into the image in place. --local uses an existing ~/.coderai, --map brings in model directories, --host-network lets it find the other boxes over mDNS (Settings → Cluster).

Windows

Docker Desktop + WSL2, NVIDIA

CoderAI-Setup-<version>.exe (download 0.2.21, sha256; also on the GitHub release) installs the coderai launcher and sets up WSL2 and Docker Desktop; or, in an elevated PowerShell:

irm https://raw.githubusercontent.com/nextime/coderai/master/packaging/windows/install-coderai.ps1 | iex

Then coderai, coderai -Stop, coderai -Upgrade, coderai -DataDir D:\coderai. Everything except the AMD/Intel Vulkan path, which WSL2 does not expose.

Signed

Verify before you run

cosign verify --key https://raw.githubusercontent.com/nextime/coderai/master/packaging/cosign.pub \
       ghcr.io/nextime/coderai:latest

Every published image carries a cosign signature; the public key is in the repo, the private half never leaves the release machine.

Offline

Move the image yourself

Pull it where there is a connection, docker save ghcr.io/nextime/coderai:0.2.21 | gzip > coderai.tar.gz, and docker load it on the offline machine.

A capability image on a machine you own is a host backend or a cluster node exactly like a pod would be: docker run -d --gpus all -p 8000:8000 -e CODERAI_API_TOKEN=… ghcr.io/nextime/coderai-images:latest, then name it on the model page or in Settings → Cluster — or add --network host -e CODERAI_CLUSTER_TOKEN=… -e CODERAI_DISCOVERY=1 and let it find the head by itself. Full list and details: docs/install-from-packages.md.

Prefer containers?

Use the published image above (docker pull ghcr.io/nextime/coderai:latest). The older coderai-docker-dist.tar bundle is still there for reference: it is the ready-to-use Docker install bundle. It includes install.sh, the coderai-docker run wrapper, coderai-dist.tar.gz containing the prebuilt Docker image, and README files. Run install.sh, then start it with coderai-docker; use it when you want immutable app code with persistent /config, /models, and /cache mounts.

Install with Docker / OCI

Platform build commands

Linux

./build.sh all      # all backends
./build.sh nvidia   # NVIDIA CUDA only
./build.sh vulkan   # AMD/Intel Vulkan path

macOS

./osxbuild.sh all
source venv_osx_all/bin/activate
python coderai

macOS uses Metal/MPS paths where available; CUDA is not a general macOS route.

Windows PowerShell

.\build.ps1 -Backend all
.\venv_win_all\Scripts\Activate.ps1
python coderai

Prerequisites and packaging

Python 3.8+.
NVIDIA: CUDA toolkit / matching driver support; CUDA 11.8+ is recommended in the README.
AMD/Intel: Vulkan drivers and SDK.
CPU-only: no GPU driver required, but heavy media generation will be slow.

./build.sh all --package
./osxbuild.sh all --package
.\build.ps1 -Backend all -Package

Packaging outputs include Linux dist-package/coderai, macOS dist-package/coderai plus CoderAI.app, and Windows dist-package/coderai.exe. Packaged builds still need compatible external GPU/runtime drivers.

Access points and local files

Path	Purpose
`/admin`	Admin dashboard, users, tokens, settings, model management.
`/chat`	Web Studio for generation tasks.
`/v1/*`	OpenAI-compatible and native multimodal API routes.
`/docs`, `/redoc`	Interactive FastAPI documentation and schema browser.

Runtime state is configuration-first under ~/.coderai/, including config.json, models.json, auth.json, pipelines.json, archive paths, and generated secrets.