Last verified

Anomaly · opencode

opencode: a coding agent that sells you nothing

Free software with fifty-two ways to attach a model — and a vendor whose own price claim we could check.

opencode is an open-source coding agent for the terminal, made by Anomaly and licensed MIT. The client is free and always will be; what costs money is the model you attach, through any of fifty-two providers, your own hardware, or the vendor's own gateway. We have not run it ourselves — everything here is read from its documentation and repository, dated and linked.

Cheapest way in Free — you pay the model
Free tier Yes
Ways to buy it 7
Oldest check 2026-08-28
§ 01 / WHAT IT IS

What it is.

Start with the identity, because three different things answer to this name.

opencode is an AI coding agent that runs in your terminal, with a desktop application and editor extensions alongside it. It reads a codebase, answers questions about it, and edits files and runs commands when you let it. The client is MIT-licensed, the repository is anomalyco/opencode, and the company behind it is Anomaly. At the time of writing the project reports 202,000 stars and 950 contributors, and the current release is v1.18.23, published three days before this page.

The name is the first thing to get straight, and the vendor knows it: its own README asks other projects using "opencode" in their name to state in their README that they are not built by this team. A separate repository under a similar name ranks alongside it in search, and the word turns up in unrelated tooling. If you are comparing notes with someone, compare repository URLs rather than names.

What makes it structurally different from the other agents in this section is that there is no product to buy. Claude Code is sold inside a Claude subscription; Codex is sold inside a ChatGPT plan and metered in credits. opencode is given away, and the vendor instead sells model access two ways alongside it — a pay-as-you-go gateway and a low-cost subscription to open-weight models — while documenting fifty-two other providers you could use instead, including running the model on your own machine. Its stated position is no lock-in, in both directions: use its gateway with a different agent, or use a different provider with its agent.

That claim is the sort we can actually test rather than repeat, so we did. The vendor says its gateway sells models at cost and adds only card processing. We took its published rate card and joined it to the twenty-five models we price independently from the labs' own pages: twenty-three matched to the cent, and the two that did not were cheaper on the gateway, not dearer. One of those carries a dated discount note on their page; the other carries no note at all. We have not run the tool itself — there are no test runs of ours behind this page, and where a claim needs evidence we do not have, it says so.

§ 02 / WHAT IT COSTS

What it costs.

Full pricing →

Nothing, and then whatever the model costs. There is no paid edition of the client and no tier to upgrade to, so every route on the pricing page buys a model rather than the program. Point it at a model on your own hardware and the bill is genuinely zero. Point it at a key and you pay the lab directly, unmarked-up. Or take one of the vendor's own two routes — a per-token gateway, or a subscription whose ceiling is unusual enough to be worth reading about: it is denominated in dollars of usage at a published rate, rather than in messages, credits or an unexplained window.

Route Bought as Price
Local model Own hardware Your hardware
Your own API key API key At model rates
OpenCode Zen API key At model rates
OpenCode Go Subscription $10/mo
ChatGPT Plus Subscription $20/mo
GitHub Copilot Pro Per seat $10/seat/mo
Enterprise Per seat Contact sales
§ 03 / WHEN NOT TO

Who it is wrong for.

Structural reasons rather than complaints about quality — cases where no amount of configuration makes this the right tool.

You want to run it on a Claude Pro or Max subscription This is the single most common thing readers try, and the documentation answers it twice, differently, on one page. The installation steps for the Anthropic provider describe selecting a Claude Pro/Max option and authenticating in a browser. An information box further down the same section says plugins that do this exist, that Anthropic explicitly prohibits it, and that opencode stopped shipping them at version 1.3.0 — while the current release is v1.18.23, so the unbundling is long done and the steps above are stale. The same box names what does work with no setup: ChatGPT Plus, GitHub Copilot and GitLab Duo. We publish the contradiction rather than resolving it, because resolving it would mean picking one of the vendor's sentences over the other. If a Claude subscription is what you hold, Claude Code is the tool it was sold to run.
You want one predictable bill and a frontier model You can have either, not both. The vendor's subscription is deliberately open-weight only — its model list is Qwen, GLM, Kimi, DeepSeek, MiniMax and friends — and reaching a frontier model means the per-token route, where the bill is proportional to use and has no ceiling. If a fixed monthly number covering the best available model is the requirement, that is what a subscription-first tool sells and this one does not.
"Free" has to mean your prompts stay yours The client is free in the way that costs nothing at all. The vendor's free hosted models are free in a different way: each is offered at no charge for a limited period while the lab behind it collects feedback, and the documentation is explicit that the data may be used to improve the model. Two of them go further and say in the docs not to submit personal or confidential data; another trades heavily discounted pricing for permission to train on your prompts and completions. Free as in beer, free as in a research programme, and free as in you own the hardware are three different offers here — the pricing page separates them. If none is acceptable, the answer is a local model or a paid key.
You want the tool to decide which model to use Provider-agnostic is the whole design, and the cost of it is that the choice is yours. Fifty-two providers are documented, each with its own auth, quality and availability — and the vendor built its curated gateway precisely because it decided that choosing well is hard, saying that with a router you can never be sure you are getting the best version of the model you asked for. That gateway is an opinion you can buy, but it is optional by design. A tool that ships one blessed model and hides the question is less flexible and, for some teams, correct.
You are on Windows and will not use WSL The installation documentation recommends WSL for Windows users for best compatibility. That is a small sentence with real consequences on a locked-down corporate laptop, where enabling a Linux subsystem is a request to somebody else rather than a decision you make. Worth resolving before you plan around the tool rather than after.
§ 04 / HOW IT COMPARES

Nothing to set it beside, yet.

All comparisons →
!
No head-to-head covers this tool yet. Head-to-head comparisons live at /compare/, where each pair is researched and dated on its own terms, and nobody has yet written one that includes this tool. We will not assemble a pair here to fill the slot — a second page on the same match-up competes with the real one for the same search, which is a problem we would be creating for ourselves. The pairs worth writing are against the other terminal agents rather than the editor-shaped tools.
§ 05 / ALTERNATIVES

If not this.

See the list →

Seven tools, grouped by the reason you would actually leave — wanting a subscription instead of a meter, wanting an editor instead of a terminal, wanting a different open-source licence, or wanting somebody accountable. Each carries its own published entry price, read on the day of writing.

Frequently asked.

What people search next.

Q · 01 What is opencode? +
An open-source AI coding agent that runs in your terminal, with a desktop app and editor extensions as well. It reads and edits a codebase, runs commands, and works against whichever model you connect — fifty-two providers are documented, from the frontier labs to a model running on your own machine. It is made by a company called Anomaly and licensed MIT.
Q · 02 Is opencode free? +
The client is, permanently: MIT-licensed, with no paid edition above it. What costs money is the model you attach, and that can also be free — either a model running on your own hardware, or one of the six the vendor's gateway does not charge for. Those hosted free models come with a condition worth reading: your prompts may be used to improve them.
Q · 03 Which opencode is this? +
The one at github.com/anomalyco/opencode, from the company Anomaly. Other projects use the same word — the vendor's own README asks them to say in their READMEs that they are not from this team — and at least one similarly named repository appears alongside it in search results. Comparing repository URLs is the only reliable way to be sure you and someone else mean the same tool.
Q · 04 Does the vendor take a cut of what you spend on models? +
Not on your own key — it is not in the billing path, so whatever the lab charges is what you pay. On its own gateway it says it sells at cost and adds only card processing, which we checked against the labs' published prices across twenty-five models: twenty-three matched exactly, and the other two were cheaper on the gateway than on the lab's own list.
Q · 05 Can I use it at work? +
The licence permits it, and there is an enterprise arrangement with a central configuration that can integrate SSO and restrict the tool to your own internal model gateway, disabling every other provider. Two things to check first: the vendor documents that it stores no code or context data but that the optional share feature sends conversations to its CDN, and it recommends disabling that for a trial. Enterprise pricing is per seat and unpublished, with the notable term that tokens are not charged if the gateway is yours.
Q · 06 Have you tested it yourselves? +
No. This page is built from the vendor's documentation, its repository, its terms of use and our own price snapshot — each read and dated on 28 August 2026. Where we say something is verified, that means we read the primary source, not that we ran the tool. The one claim we did test independently is the pricing one, because it could be checked against data we already hold.