ChatDev

The trait that decides it is the starting point. This produces software from a description, with agents in named roles handing work along, and it is genuinely good at going from an empty directory to something that runs. That is a demo you will enjoy and a workflow you will use roughly twice.

Real work starts from a repository with history, conventions and tests, and nothing here is designed for that. Pick OpenHands when the codebase exists, and keep this for teaching, for experiments, and for the afternoon you want to see a virtual company argue with itself.

The risk is a pivot mid-citation. The 1.x line that made this famous now lives on a legacy branch, and the current product is a configuration console with an SDK beside it. Three years of write-ups, tutorials and academic references describe behaviour that is no longer the default, and a reader has no way to tell which one a given article meant.

Check the branch before reproducing anything you read. What it does right: the classic line was moved rather than deleted, so the code behind the published work stays available to anyone checking it.

The design deserves credit for being explicit and criticism for being rigid. Phases run in a fixed order, each one a dialogue between agents holding assigned roles, and the artefacts of one phase become the input to the next. A specification error therefore propagates forward with no mechanism to send it back, which is the same weakness the human process it imitates has always had.

Testing appears as a phase in that sequence. Nothing in the record documents an executed suite gating progression, and no benchmark is reported, so the pipeline's yield is undescribed.

Thirty-four thousand stars is one of the larger numbers in this class and it has produced no commercial entity, which tells you the project's currency is citations rather than customers. Research groups optimise for papers and for the next grant, and both survive perfectly well without a single user paying anything.

The zero-code console in the current version is the shape of a product looking for a buyer, and no pricing has followed it. Likely path: continued academic output with occasional product gestures. Position: read the papers, do not model a vendor relationship that does not exist.

Standing it up means a Python backend and a Node front end per engineer, which is two toolchains our platform team would then be maintaining for a tool with no unattended mode and therefore no place in our pipelines. Access control does not exist because there is no shared deployment to control.

The supplier is a research group, so there is no agreement, no security contact and no commitment to anything. I have no objection to an engineer running it for a week and reporting back. As something sixty people depend on, not yet.

The model configuration is refreshingly blunt: your own key and your own base URL, which means anything speaking the common API shape works, including whatever I am running locally behind a proxy. Standing it up is uv sync, an npm install and make dev, and I had it running before I had finished reading the README.

No MCP on either side, so tools are whatever the codebase defines and extending it means editing the codebase. Under Apache-2.0 that is fine by me. This is a project I would fork to learn from rather than to deploy, and it is pleasant to read.