OpenCode is a free, open-source AI coding agent built by Anomaly, the team behind SST. It runs inside your terminal, desktop app, browser, or IDE, and connects to more than 75 model providers so that a language model can act as a full software engineer on your machine. It reads your codebase, edits files, and runs commands without sending your source to anyone’s server. Because the program is MIT-licensed and model-agnostic, the agent itself costs nothing. You pay only for the language model you point it at, and when you point it at a free or locally hosted model, the entire stack can run for zero dollars.
That “free, but optional” framing is the whole point of OpenCode, and it is a large reason it has become the most-starred coding agent on GitHub. Below is what the program actually is, how it works, and exactly what it costs, including the free and paid paths.
What Is OpenCode?
OpenCode is an open-source AI coding agent. It is a program, not a hosted service, that lives on your machine and orchestrates a language model so that model can behave like a working developer.
- Developer. Anomaly, the same team behind the SST (Serverless Stack Tooling) framework.
- License. MIT. The agent, the terminal UI, and the IDE extensions are all open source, with a very large contributor base and hundreds of thousands of GitHub stars.
- Where it runs. Your terminal via a polished TUI, a desktop app, a browser app, or IDE extensions.
- What it does. Reads and edits files, runs shell commands, uses LSP diagnostics and renames, and manages multi-session agent workflows, all on the device you control.
The defining trait is that OpenCode is model-agnostic. It supports 75+ providers through the Models.dev catalog and the AI SDK, so you can switch between frontier models, open-weight models, local models, or an existing coding subscription with a single config change.
How OpenCode Works
OpenCode uses a client-server architecture. The part you install on your machine talks to a language model through a provider of your choice, then carries out the model’s instructions by editing files and running commands locally. A few architectural details matter for why people choose it.
- Bring your own key. You connect a provider with a single command, and your API key is stored locally. No OpenCode account is required to run the agent.
- LSP integration. It hooks into language servers for diagnostics, completion, and rename actions, so the agent edits code the way an IDE does instead of guessing at raw text.
- Multi-session by design. Several agents can work on the same project in parallel without conflicting, which suits larger codebases and parallel tasks.
- Zero data retention. OpenCode stores no code or context on any vendor server, which is why regulated and air-gapped teams can run it entirely on their own infrastructure.
Is OpenCode Free? The Pricing Reality
This is where the name trips people up, and where the free-versus-subscription distinction matters. OpenCode is a free program with an optional subscription, and the two are not the same thing.
The program is free
The MIT-licensed agent itself costs $0. You install it, point it at any model provider you already use, and pay only for that model’s usage. Nothing in OpenCode charges you for the orchestrator, and you do not need an account to run it.
The subscription is optional
Anomaly also sells an optional subscription called OpenCode Go, priced around $10/month with the first month at $5. Go gives you reliable access to a curated set of capable open-source models with generous limits, so you do not have to source and meter your own keys. It is a convenience layer, not a requirement. You can use Go with OpenCode or with other agents, and you can cancel at any time.
An optional managed gateway
There is also OpenCode Zen, an optional pay-as-you-go model gateway that serves curated, benchmarked models tuned for agentic coding and bills per million tokens. Per-seat enterprise plans exist for managed deployments. All of these are add-ons that sit on top of the free core, not replacements for it.
Truly free with a free or local model
Because OpenCode is model-agnostic, you can remove the model cost entirely.
- Point it at a free model tier, such as the free models on OpenRouter or currently-free options like the Agnes 3.0 Flash model.
- Run a locally hosted model through Ollama, LM Studio, or llama.cpp. In that setup, the agent, the tooling, and the inference all run on your own hardware, and the LLM cost is effectively zero.
The practical takeaway is that the agent is free, the subscription is optional, and the model is whatever you choose, including a free one. OpenRouter is a good place to browse free and cheap models, and opencode.ai documents every connection path.
| Option | What it costs | What you get |
|---|---|---|
| The agent itself | Free forever, MIT-licensed | Full TUI, desktop, browser, and IDE surfaces; no account required |
| Bring your own key | Pay only your model provider | 75+ providers including frontier, open-weight, and existing subscriptions |
| Free or local model | $0 | OpenRouter free tiers, currently-free models, or your own hardware |
| OpenCode Go (optional) | $10/month (first month $5) | Managed access to curated open-source coding models, generous limits |
| OpenCode Zen (optional) | Pay per million tokens | Benchmarked model gateway curated for agentic coding |
OpenCode Versus Closed Agents
OpenCode’s biggest structural advantage over closed agents such as Claude Code is that it has no vendor lock-in. You are not tied to one company’s model, one pricing tier, or one roadmap.
- Flexibility. Closed agents are usually locked to a single model family. OpenCode works with 75+ providers and local models.
- Self-hosting. Because nothing is retained server-side, teams that cannot send source to a third party can still run a capable agent.
- Cost control. A closed agent’s subscription bundles model access you cannot change. With OpenCode, you pick the model and pay only for it, or nothing at all with a free or local model.
Even closed agents are powerful, and it helps to understand why. The Claude Code architecture piece on this site explains how orchestration and infrastructure, not just raw model intelligence, drive their performance. That same orchestration layer is something you can run for free with OpenCode and a model you already control. And if you do want a closed agent but are watching costs, cheaper alternatives such as Meta Muse Code exist alongside the big names. If you want to run several agents at once, the Herdr open-source terminal pairs well with OpenCode’s multi-session model.
Is OpenCode Safe?
OpenCode stores no source or context data server-side, which is a strong default for confidential work. Keep in mind, though, that it is a local binary with broad filesystem and shell access, and it has a disclosed 2026 security history that self-hosting teams should review. On shared or untrusted networks, run its local server with extra hardening, and treat any agent that can read files and execute commands as powerful software that deserves the same care you would give a privileged process.
How to Get Started With OpenCode
Getting set up takes minutes. The full project is on GitHub, and the steps are straightforward.
- Install the agent. Use the macOS/Linux shell script, the npm/Bun package, or the Windows one-liner for your system.
- Launch the terminal UI. Run opencode to start the TUI.
- Connect a model. Use a /connect command to store a key locally. Choose any of the 75+ providers, a free model, an existing ChatGPT or Copilot subscription, or a local model.
- Start a session. Point OpenCode at a project and begin. The agent reads, edits, and runs against your code from there.
Who Is OpenCode For?
OpenCode is a good fit for several kinds of developers, and an awkward fit for others.
- Hobbyists and students on a budget. Pair the free agent with a free model tier or a small local model and the entire AI coding stack costs nothing. This is the lowest-cost way to learn how agentic development actually works.
- Privacy-conscious teams. Zero server-side data retention means regulated industries, air-gapped networks, and anyone under NDA can run a capable agent without shipping source code to a vendor. It is the same self-hosted philosophy behind projects like OpenClaw, the open-source AI agent you run on your own infrastructure.
- Model tinkerers. If you like benchmarking different models against the same codebase, switching providers with a single config change is far faster than migrating between locked-down tools.
- Developers who already pay for a coding subscription. ChatGPT Plus or Pro, GitHub Copilot, and GitLab Duo logins work with zero extra setup, so an existing seat becomes the model budget at no added cost.
It is less ideal if you want a single vendor to handle everything with a traditional support contract, or if you are uncomfortable managing your own API keys and configuration. In those cases, a closed, fully-managed agent may be the better starting point, and you can switch to OpenCode later without lock-in once you know what you need.
Frequently Asked Questions
Is OpenCode really free?
Yes. The agent is free and MIT-licensed. You pay for the model you use, or nothing at all if you use a free or local model. The $10/month Go subscription and the Zen gateway are both optional add-ons, not requirements.
Can OpenCode use ChatGPT or other subscriptions?
Yes. It can connect to ChatGPT Plus or Pro, GitHub Copilot, and GitLab Duo logins with zero extra setup, so a seat you already pay for becomes the model budget.
Is OpenCode better than Claude Code?
It depends on what you are optimizing for. OpenCode does not bundle its own frontier model, so on raw model intelligence the answer depends entirely on the model you connect. Claude Code remains the benchmark for pure capability with Claude models, and its subscription bundles that power in one place. OpenCode’s advantage is choice and control: pair it with any strong model, or a free one, and you get that model’s power with no lock-in, no server-side data retention, and a fully free harness. Many developers use both, reaching for a closed agent when raw capability matters most and OpenCode when cost, privacy, or provider flexibility wins.
What is OpenCode Go?
Go is OpenCode’s optional $10/month subscription, with the first month at $5. It provides reliable access to a curated set of open-source coding models with generous limits, and you can cancel at any time.
