There's a quiet decision buried in every AI coding tool: who owns your workflow? Most of the popular ones answer "we do" — the agent and the model ship together, so your day-to-day depends on one company's pricing, uptime, and roadmap. OpenCode answers differently, and this month that answer reached scale: it crossed roughly 7.5 million monthly developers and 184K GitHub stars, shipping v1.17.18 on July 9. For founders, the interesting thing isn't the star count — it's the design choice underneath it.

What it is: OpenCode is a free, MIT-licensed, terminal-first AI coding agent from Anomaly (the team formerly known as SST). Its defining feature is that it doesn't own the model: you bring your own key and point it at any of 75+ providers, or at a model you self-host.

What it is#

OpenCode runs in your terminal and gives you two built-in agents: a build agent with full access to read and change your code, and a plan agent that's read-only, for analysis you don't want touching files. You switch between them with a keystroke. Under the hood it wires into your project the way an engineer expects: LSP integration for 20+ languages (so the model gets real compiler diagnostics, not guesses), and MCP support to connect GitHub, Postgres, Slack, and any custom tool you expose.

The load-bearing difference from most coding agents is the model relationship. OpenCode is model-agnostic by design — 75+ providers, your own keys — and because it's MIT and self-hostable, you can run it locally or even air-gapped, which is the difference between "maybe someday, for regulated teams" and "today." (We put it side-by-side with the incumbent in OpenCode vs. Claude Code; this piece is about the tool on its own terms.)

Who it's for#

If you're an engineer or a technical founder who wants a capable agent with control — pick the best or cheapest model today, switch tomorrow, keep your source on your own machine — this is squarely aimed at you. It's also the right call for anyone with a compliance reason to keep code off third-party clouds: self-hosting and air-gapping are first-class, not an afterthought.

Who it isn't for: a non-technical founder who wants to click around a codebase from a friendly web UI. OpenCode is terminal-first, and its power lives in the TUI, the MCP connections, and the LSP feedback loop — none of which land if you're not comfortable in a shell. For that reader, a hosted point-and-click tool is a better fit, at the cost of the portability OpenCode is built around.

How to start#

One command:

curl -fsSL https://opencode.ai/install | bash
# or: npm i -g opencode-ai@latest   (Homebrew and Scoop also work)

Then set an API key for whatever provider you want — a frontier hosted model, a many-models router, or a model you self-host on your own hardware — and run opencode inside your project. You'll drop into the TUI with the build/plan agents ready. Because it speaks to 75+ providers, the same install works whether you're on a frontier model this week and a cheaper one next.

What it costs#

The tool is free and open source. The only thing you pay for is the model — OpenCode is bring-your-own-key, so your cost is whatever your chosen provider charges for tokens. Versus a bundled single-vendor subscription, that's the trade: you manage the model bill yourself, and in exchange nothing about your workflow is tied to one company's pricing. If you already hold provider credits, running OpenCode is effectively free.

The honest catch#

Three of them. You supply and pay for the model — there's no bundled model, which is the entire point but also a setup step. It's terminal-first, so the value doesn't reach a non-technical user. And "bring your own everything" is a little more configuration than a fully hosted tool — the reward being exactly the thing the run-anywhere week is about: a core dev workflow that no single vendor can reprice out from under you.

The star count is a proxy. The real signal is 7.5M developers voting that their coding agent shouldn't come welded to one model company.