OpenDevin’s Early Coding Agent: A 2024 Installation and Demo

Matthew Berman introduces OpenDevin in March 2024 after Cognition’s Devin demonstration drew attention. He likes Devin’s combination of dialogue, terminal, browser and editor, and presents OpenDevin as an independent open-source project inspired by that interface, not as the same product or its source code.

The project later became known as OpenHands, and its setup evolved after this recording. Berman’s video description also discloses that he is an investor in CrewAI. His walkthrough records a very early development version, with the limitations he encountered.

Berman asks the agent to write a simple Python calculator. The workspace displays dialogue alongside terminal, planner, code editor and browser tabs. He uses a LangChain agent and a GPT-4 model; the interface also lists a CodeAct option, which he says he had not tested.

The agent inspects files, works on the calculator and runs commands. Berman follows up by asking for an HTML interface and shows a basic calculator that he says works. The on-screen agent log also says it could not directly test that interface via an HTTP server and instead reviewed the code, so the example does not establish fully autonomous end-to-end verification.

Berman runs the interface locally while using a hosted GPT-4 model, and says another compatible model could be configured. He stresses that the project is only days old in the video and that features remain buggy, even though he finds the pace of development impressive.

He notes the project’s early GitHub attention and mentions another Devin-inspired project, Devika, which he had not been able to get working. He then begins the OpenDevin installation, warning that some of the issues he encountered involved Python packages and environments.

In Visual Studio Code, he opens a terminal, clones the repository and opens its folder. He creates and activates a Conda environment named od with Python 3.10. These instructions describe the early version shown, not a tested installation procedure for the current project.

He checks that Docker is installed and running with docker ps, installs Docker Desktop if necessary and pulls the project’s sandbox image. He values the containerized execution environment, although using Docker alone is not a guarantee that arbitrary agent-generated code is safe.

For the hosted-model example, Berman creates an OpenAI API key and exports it as an environment variable; he says he will revoke the demonstration key. He notes that shell exports must be set again after a terminal restart and mentions a local environment file as an alternative. Credentials should remain private rather than being reproduced in the walkthrough.

He sets a workspace directory and installs Python requirements using the active environment’s interpreter. To diagnose dependency problems on his machine, he checks which Python is in use. His commands and paths reflect that particular 2024 setup.

Berman explains that he had needed Rust for an orjson dependency and restarted the terminal after installing it. He subsequently removed and reinstalled orjson using a binary-only pip option that worked on his Apple Silicon machine. He cautions that other machines may encounter different errors.

He starts the backend with Uvicorn on localhost port 3000. It appears to hang, and a later attempt starts successfully; he does not know why. Although he describes Ctrl-Z as force quitting, the usual shell behavior is to suspend a foreground process, not terminate it.

In another terminal he checks the working directory and environment, enters the frontend folder, installs Node packages with npm and starts the frontend on port 3001. He opens the local page and selects GPT-4 for a simple hello-world website task.

The agent creates an HTML file and starts a server in his example. Berman observes that the planner seems unreliable and that the browser view has problems; the backend logs show an error and exit. He regards these as signs of a promising but still buggy early project.

He describes changing model-provider settings, including an API key and model name for Claude, or a base URL for a locally served OpenAI-compatible endpoint. He mentions LM Studio and Ollama as possible local servers. Compatibility, model quality and configuration depend on the project version and endpoint.

Berman says Devin’s polished interface impressed him and that OpenDevin’s early progress is encouraging. He invites users to try it and report issues, while his limited calculator and hello-world examples do not demonstrate reliable autonomous development of arbitrary applications.