codeeli.com

What it is

Codeeli is a self-hosted code generation web app. Its goal is not to be a chatbot or a general-purpose agent, but to use local 4B–8B small models to generate “small projects that can run.” After installing it on Linux or macOS, users can connect Ollama, OpenAI-compatible, or LiteLLM models, and write generated files directly into a local workspace.

Core capabilities

Its key design is “plan-then-stream”: first, it asks the model to return a strict JSON file list, including filenames and a one-line task description; then it generates each file one by one, using existing files as context. If the plan is not valid JSON, it is discarded and retried up to three times, with no fuzzy repair. Prompt templates and the recipe library are stored in the database and can be edited through the UI, allowing users to tune behavior for different models.

Pricing and deployment

The main text presents Codeeli as an open-source project, with GitHub source code and a curl installation script available. No paid plans are mentioned. It does not provide a cloud demo; the official explanation is that the app writes local files and executes shell commands, so putting it on a public cloud would introduce security risks or fail to simulate the real experience. Therefore, the way to try it is to install it locally and connect your own model.

Pros and cons

Its strength is its clear positioning: rather than asking a small model to generate hundreds of lines in one large response, it breaks the task into planning and file-level generation, which is more robust from an engineering perspective. Self-hosting also keeps the code, model, and workspace on the user’s own machine, offering better data control. The downside is that the barrier to entry is higher than SaaS: it requires git, Python 3.10+, and a local model environment. It is also not a complex project agent, and there is limited information about online hosting, team collaboration, permission isolation, or security sandboxing.

Who it is for and access from China

It is suitable for developers familiar with local LLMs who want to use small models to generate prototypes such as a single index.html or a tiny Python CLI. The main text does not specify access conditions from China. If users need to pull source code and installation scripts from GitHub, the actual experience may be affected by the network environment. Alternatives include Continue, Aider, OpenHands, Cursor, GitHub Copilot, or using Ollama directly with a local code model.