OpenCode: The Open-Source Coding Agent You Run on Your Own Machine

Most AI coding agents ask you to pick a side: pay a monthly subscription and use their model, or bring your own key and figure out the rest yourself. OpenCode refuses that trade. It is a free, MIT-licensed coding agent that runs on your machine, in your terminal, and it will happily work with almost any model you point it at — Claude, GPT, Gemini, DeepSeek, or a model running on your own hardware. The numbers explain why it keeps showing up in roundups of free coding agents this month. The repository at github.com/anomalyco/opencode has 222,599 stars as of October 28, 2026, with 29,798 forks and more than 21,500 commits since the project was created on April 30, 2025. The project's own site claims over 16 million developers use it every month — treat that as the project's own figure, not an independent count — but the star count is verifiable on the repo page right now. This is a review of OpenCode as its own tool: what it actually installs, what the free tier really buys, where the money goes when you do pay, and the catches that matter before you hand it your codebase. ## What OpenCode actually is OpenCode is an open-source agent that writes, edits, and runs code from your terminal. Version 2.0, released in October 2026, runs one background service that powers the terminal interface, a desktop app (currently in beta), and a web app, so sessions can keep running while you switch between them. It is also available as an IDE extension. Inside a session, two built-in agents cover the two modes every developer actually uses. The **build** agent is the default: full access, it edits files and runs commands. The **plan** agent is read-only — it denies file edits by default and asks permission before running shell commands — which is the right mode for exploring an unfamiliar codebase or agreeing on an approach before anything changes. You switch between them with the Tab key. A **general** subagent handles complex searches and multi-step tasks, and you can call it directly with `@general`. The workflow the docs teach is deliberately boring, which is a compliment. You open a project folder, run `opencode`, and run `/init` once so the agent analyses the project and writes an `AGENTS.md` file describing its structure and coding patterns. From then on you describe changes the way you would brief a junior developer, ask for a plan first on anything risky, and switch back to build mode when the plan looks right. If a change goes wrong, `/undo` reverts it and shows your original message again; `/redo` re-applies it. Three features separate it from a thin terminal wrapper. First, **multi-session**: you can run several agents in parallel on the same project. Second, **subagents that run in the background** in version 2, so a long task does not block your terminal. Third, **Code Mode**, which chains MCP tools into a single script — the project's way of letting one prompt drive several tools in sequence. ## How it differs from Claude Code The comparison everyone asks for is Claude Code, and the difference is structural, not cosmetic. Claude Code is Anthropic's agent: it runs Anthropic's models, it ships with Claude Pro at $20 a month (or Max from $100, or API billing), and there is no free tier. OpenCode is model-agnostic: the project lists more than 75 providers through Models.dev, including Claude, GPT, Gemini, and local models through Ollama, LM Studio, or vLLM. That shows up in four practical ways: - **No subscription is required.** The app is free and open source under the MIT licence. You can use it forever without paying OpenCode anything. - **You can bring subscriptions you already pay for.** OpenCode lets you sign in with a ChatGPT Plus or Pro plan, or a GitHub Copilot account, and use that plan's models inside the agent. Independent coverage also lists GitLab Duo and SuperGrok as usable the same way. - **Claude is the exception, not the rule.** There is no Claude Pro or Max sign-in in OpenCode — independent reporting notes it has been gone since version 1.3.0 in March 2026, and Anthropic's terms bar other developers from offering one. Claude models still work in OpenCode, but only through your own Anthropic API key. The homepage's "including Claude" refers to models, not to signing in with a Claude subscription. - **Your code stays local by design.** The project states it does not store your code or context data, which is why it positions itself for privacy-sensitive environments. One exception matters: your data is stored when you use OpenCode's own free models — see the data-terms table below. | | OpenCode | Claude Code | Codex | | --- | --- | --- | --- | | Licence | MIT, open source | Closed source | Closed source | | App cost | $0 | $0 app, but no free tier for the model | Included with ChatGPT plans | | Models | 75+ providers, any model, local models | Anthropic Claude only | OpenAI GPT models | | Bring your own subscription | ChatGPT Plus/Pro, GitHub Copilot | Claude Pro $20/mo, Max from $100 | ChatGPT Plus $20/mo, Pro from $100 | | API access | Yes — API keys for any provider, MCP support, Zen and Go gateways | Yes — Anthropic API billing | Yes — OpenAI API billing | | Runs on | Terminal, desktop app (beta), web app, IDE extension | Terminal, IDE | Terminal, IDE, ChatGPT app | Prices for Claude Code are from our own Claude Code explainer; Codex pricing is from OpenAI's published plan pages. For the wider field, see our roundup of the 7 best AI coding agents in 2026. ## What it costs — and the fine print The honest summary is one sentence: the agent is free, the model usually is not. There are four ways to pay for the model, and the fine print differs for each. | Route | Cost | The catch | | --- | --- | --- | | OpenCode app (CLI, desktop, IDE) | $0, MIT licence | None on the app itself | | Free models | $0, no account needed | Rotating lineup, "limited time", unpublished usage limit, data terms vary by model | | OpenCode Zen (pay-as-you-go) | $20 top-up + $1.23 card processing fee | Per-request billing at provider rates, zero markup | | OpenCode Go | $10/month; Go Plus $40/month | Plan usage limits, single-member workspace | | Your own provider key or subscription | Your provider's rates or plan | That plan's limits and overage | **Zen** is OpenCode's curated gateway: a handpicked set of models the team says it has tested and benchmarked specifically for coding agents. You add a $20 pay-as-you-go balance — plus a $1.23 card processing fee per top-up — and you are billed per request at the provider's own price, with zero markup. You can set monthly spend limits, disable billing any time and keep using your remaining balance, and Zen works with any agent, not just OpenCode. Refunds are available if the charge was made within the last 14 days and you have not used the credits. All Zen models are hosted in the US and EU, and providers follow a zero-retention policy and do not use your data for training, with listed exceptions. **Go** is the flat subscription for people who want predictable pricing on open models. It costs $10 a month, Go Plus costs $40, and both can be cancelled any time. The lineup is curated and tested for agentic coding, and as of October 28 it includes Kimi K3.1 Pro, Gemini 4 Argon, Claude Haiku 5.5, GPT 6 Luna, MiniMax M3.1 Flash, DeepSeek V4.1 Flash, Muse Spark 1.3 Contributor, Grok 4.7, and about twenty more. Limits are enforced across rolling windows: 20% of the monthly allowance per 5 hours, 50% per week, 100% per month, and the request counts vary sharply by model — Kimi K3.1 Pro gets an estimated 147 requests per 5 hours on Go, while DeepSeek V4.1 Flash gets about 34,670, because the allowance is priced in dollars behind the scenes. If you need more usage, you top up credit in your account. The same 14-day refund rule applies if you have not used your allowance. Note the pricing trap that catches people comparing the two: Zen and Go are different products. Zen is pay-as-you-go with no monthly fee; Go is a flat subscription. You do not need either — the app works with your own keys and with the free models — but if you want OpenCode's curated models without managing provider accounts, pick Zen for sporadic use and Go for steady daily use. ## The free models, and the catch in their data terms The free tier is the reason most people install OpenCode, and it is also the part that needs the closest reading. The app ships with a set of free models you can use without creating an account. The lineup rotates, every model is described as free "for a limited time" with no end date published, and there is an unpublished usage limit — independent coverage in late October reported hitting a "free usage limit reached" message with no documented cap. One more wrinkle: the official models page at opencode.ai/models currently returns a 404, so the live list is checked inside the app itself with `opencode models`. The table below is a snapshot from independent coverage dated October 23, 2026, cross-checked against the free models named on OpenCode's own Go page. Treat it as a snapshot, not a contract. | Free model | Data terms | | --- | --- | | DeepSeek V4.1 Flash Free | Zero retention, no training on your data | | GLM-5.3-Flash Free | Zero retention, no training on your data | | LongCat 2.5 Preview Free | Zero retention, no training on your data | | Beam Free | No training, retained 30 days, not available in some regions | | Big Pickle | Stealth model; data "may be used to improve the model" | | Flint 40B Free | May be used to improve the model, retained 30 days | | Exo Free, MiMo-V2.6-Flash Free, MiMo-V2.5 Free | May be used to improve the model | | Nemotron 3 trial models | NVIDIA trial endpoints: "do not submit personal or confidential data"; use is logged | The practical rule is simple. If a model says zero retention and no training, it is safe for ordinary project code. If it says your data "may be used to improve the model", treat it like a public pastebin: fine for a side project, not for proprietary code, credentials, or anything under NDA. And the NVIDIA trial endpoints carry an explicit warning not to submit personal or confidential data at all. The free tier is genuinely free — but the price is paid in data terms, and the terms differ model by model. ## How to install it and run your first session The current install, from the project's own README, is: ``` curl -fsSL https://opencode.ai/v2/install | bash ``` Package managers work too: `npm i -g @opencode/cli@latest`, `brew install anomalyco/tap/opencode-v2` on macOS and Linux, `sudo pacman -S opencode` on Arch, plus mise, Nix, and a Docker image. The desktop app is in beta and available from opencode.ai/download, or with `brew install --cask opencode-desktop` on macOS. Remove versions older than 0.1.x before installing. Once it is installed, the first run takes about two minutes: 1. Run `/connect` in the terminal UI and pick a provider — OpenCode's own Zen, one of the free models, or any provider key you already have. Signing in to Zen takes you to opencode.ai/auth, where you add billing details and paste the API key back. 2. Change into a project folder and run `opencode`. 3. Run `/init` once so the agent writes an `AGENTS.md` describing your project. 4. Ask it to explain something first — "How is authentication handled in this codebase?" — before asking it to change anything. Press Tab to switch to plan mode for anything you want reviewed before it touches a file. Themes, keybindings, code formatters, custom commands, and the full configuration file are all documented on the project's docs site, and the whole configuration is plain files you can version-control. ## The catches, stated plainly First, the free models are a rotating promotion, not a permanent tier. The lineup changes, the usage cap is unpublished, and the official page listing them currently 404s. If your workflow depends on a specific free model, expect it to disappear one day without warning. Second, the strongest models are never free. The free tier is genuinely useful for small tasks and for trying the tool, but sustained work on a large codebase will push you to Zen, Go, or your own provider key — and at that point your cost is the model's cost, not OpenCode's. Budget for the model, not the app. Third, the Claude gap matters if you are coming from Claude Code. There is no Claude Pro or Max sign-in, so the $20 Pro plan that makes Claude Code cheap does not transfer. Claude models in OpenCode mean API billing, which is metered and can surprise you on a long session. Fourth, version 2 is new, and the upgrade has edges. The project's own docs warn that OpenCode 1 and OpenCode 2 installs can conflict, and older plugins do not run in 2.0. If you are following a tutorial written for version 1, check which version it targets before running its install commands. Fifth, this is a tool that runs code on your machine with your permissions. The plan agent exists for a reason: use it for unfamiliar codebases, review what build mode is about to do, and keep secrets out of prompts sent to any model that trains on data. ## Pros and cons **Pros:** free and MIT-licensed with no paid tier for the app itself; works with 75+ providers plus local models, so there is no lock-in; free models included with no account; you can reuse ChatGPT Plus/Pro or GitHub Copilot subscriptions you already pay for; terminal, desktop, web, and IDE surfaces share one background service; plan mode, undo/redo, and multi-session are genuinely useful safety nets; privacy-first design with no code or context stored by the app itself. **Cons:** free models rotate, carry an unpublished usage cap, and have data terms that vary by model; the strongest models always cost extra, so the real bill is the model's; no Claude Pro/Max sign-in — Claude means API billing; version 2.0 is new and older plugins and version-1 tutorials do not carry over; the most impressive usage numbers on the project's site are its own claims, not independent measurements. ## FAQ ### Is OpenCode really free? Yes, for the app. It is MIT-licensed, the source is public on GitHub, and it ships with free models that work without an account. You pay only for model usage beyond the free tier — through Zen pay-as-you-go, the Go subscription, your own provider keys, or a subscription you already hold, such as ChatGPT Plus or GitHub Copilot. ### Does OpenCode work with Claude models? Yes, through an Anthropic API key. What does not work is signing in with a Claude Pro or Max subscription — that sign-in has been unavailable since version 1.3.0, and Claude models in OpenCode are billed through the API instead. ### Which models are free? A rotating set, including DeepSeek V4.1 Flash, GLM-5.3-Flash, LongCat 2.5, Beam, Big Pickle, and Flint 40B in late October 2026. Every free model is described as free "for a limited time", the usage cap is unpublished, and the data terms differ by model — check them in the app with `opencode models` before pasting in sensitive code. ### Does OpenCode store my code? The project states it does not store your code or context data, which is the basis for its privacy-first positioning. The exception is its own free models: your data and information are stored when you use those. Paid providers follow their own retention policies, which the Zen and Go pages list model by model. ### Do I have to use the terminal? No. The terminal interface is the original surface, but version 2 added a desktop app (currently in beta) and a web app, plus an IDE extension. The terminal is still where the full feature set lives first. ## Key takeaways - OpenCode is a free, MIT-licensed coding agent with 222,599 GitHub stars as of October 28, 2026 — the app itself costs nothing, forever. - It is model-agnostic in a way Claude Code is not: 75+ providers, local models, and sign-in with ChatGPT Plus/Pro or GitHub Copilot subscriptions you already pay for. - The real cost is the model: Zen is $20 pay-as-you-go top-ups with zero markup, Go is $10 a month ($40 for Go Plus), and your own provider keys bill at provider rates. - The free models need no account, but the lineup rotates, the usage cap is unpublished, and some models may train on your prompts — read each model's data terms before using it on real code. - Install it with the v2 script from the project's README, run `/init` once per project, and use plan mode before letting it change code you have not reviewed. OpenCode's pitch is not that it has the smartest model — it does not have a model at all. Its pitch is that the agent should be yours: free, open source, running on your machine, pointed at whatever model you choose. As of October 28, 2026, with 222,599 stars and a second major version shipping this month, that pitch is finding a very large audience. Just go in knowing the free models are a rotating offer with strings attached, and budget for the model, not the app.