What is opencode?

opencode is an open source coding agent that lives in the terminal, released under the MIT license and installed on the machine of the person writing the code. What it does resembles its peers: it reads a repository, proposes and executes changes across multiple files, runs commands and tests, and connects to external tools. The difference that matters to whoever signs off is not functional but architectural, and it is independence from the model vendor. The official documentation lists dozens of providers configurable with an API key, alongside models running locally through OpenAI-compatible endpoints: the model becomes a configuration parameter rather than a contractual commitment. An organization working with one vendor today that has to switch tomorrow, over price, performance or a clause on data handling, changes a line of configuration and not the tool its developers work in. That reversibility is the property worth evaluating, not the open source label.

What it does

Operationally it is a command-line agent harness: a loop that explores the code, plans, edits and verifies, with per-project configuration, language server integration, permissions over which actions are allowed and connections to external tools. The second point that changes the corporate conversation is where the code stays. In its main form the agent runs on the developer's workstation and reads the repository from there, so the exposed surface is the call to the chosen model, not a managed service hosting the working session. If that model is a local LLM, the call never leaves the corporate network at all, and the conversation with legal about confidentiality and ownership of source code changes in nature.

When to choose it, when not to

It fits where vendor reversibility is an explicit requirement, where a policy already forbids sending source code to an outside service and where someone in house can manage models and configuration. It should be ruled out, and this is worth saying plainly, when procurement signs the decision: there is a company behind the project, Anomaly, but the distribution model is open source, and it carries no support contract with response times, no centralized per-seat license management and none of the compliance documentation a tender asks for. Self hosting and provider selection are likewise work that someone has to do and keep doing. Against its siblings it is the terminal option not tied to a single lab: Claude Code and OpenAI Codex offer one vendor to sign with, which for many organizations is an advantage rather than a limit. The right question is not which tool is better, but whether the lock-in on the model is a risk the organization wants to keep open.