Pick OpenCode if you want a free coding agent that is not tied to one vendor's models. The vendor says 16 million developers use it every month. Expect to bring your own provider keys or add a paid plan, and review the approval settings before letting it run shell commands.
What changed
- · Release OpenCode 2 available alongside the 1.x line (2.0) source
OpenCode is an MIT-licensed, open source AI coding agent that runs in a terminal, desktop app or IDE extension. It connects to more than 75 language model providers through Models.dev, including local models, and runs several agent sessions in parallel on one project. Its GitHub repository shows over 200,000 stars as of October 2026.
Maker: Anomaly · Autonomy: semi autonomous · Maturity: GA
Underlying models: Any model (75+ providers via Models.dev, including local models)
About OpenCode
OpenCode is an open source coding agent made by Anomaly. You give it a task in plain English, it reads your project, edits files and runs shell commands, and it asks before doing anything your permission settings flag. That plan-act-ask loop is what separates an agent from a chat window that only suggests code. The source is MIT licensed, and the GitHub repository lists more than 950 contributors.
The design choice that defines it is that it is not tied to one model. It connects to more than 75 providers through Models.dev, including local models, and you can sign in with an existing GitHub Copilot or ChatGPT Plus or Pro account instead of pasting keys. It loads the right language servers (LSP, the same code-intelligence protocol your editor uses) automatically, starts multiple agents side by side on the same project, and can share a link to any session for debugging. Compare that with Claude Code and Gemini CLI, which are built around their own vendors' models, and OpenAI Codex, which is built around OpenAI's.
It suits developers who already work in a terminal and want to pick the model per task, teams that need to keep code inside their own gateway, and anyone who prefers to audit the tool they hand shell access to. Cline is the closest rival if you would rather stay inside VS Code. Terminal first is the trade: there is more configuration to learn than in a one-click IDE assistant such as Cursor or Windsurf. Not sure which agent fits your stack? Smart Match can narrow it, and the coding agents directory lists the rest.
The agent itself is free. You pay your model provider directly, or you buy model access from OpenCode: Zen is a pay-as-you-go balance with a curated list of models the team has tested for coding agents, and Go is a monthly subscription for open-source models, with a higher Go Plus tier. Enterprise is priced per seat, with central config and single sign-on. The tiers and amounts are in the cost FAQ below.
The project moves quickly: the repository logs new commits daily, and OpenCode 2 now sits alongside the original 1.x line, with its own install path. That pace brings churn, so pin a version if you run it in a team. Security deserves a read before you start, because the agent can execute commands with your privileges.
Pricing
The OpenCode agent is free and MIT licensed; you pay your model provider. Zen is pay as you go: a $20 starting balance (plus card processing fee) that auto-tops up by $20 when it falls to $5, with no markup on model prices per the vendor. Go is $10 per month and Go Plus is $40 per month, both with request limits per 5 hours and monthly usage caps that vary by model. Enterprise is per seat, price on request. Checked 2026-10-18.
Key Features
- Any model, any provider: Connects to 75+ LLM providers through Models.dev, including local models, and accepts GitHub Copilot or ChatGPT Plus and Pro logins.
- Automatic LSP loading: Starts the matching language server for each project so the model sees type errors and symbols, not just raw text.
- Parallel sessions: Runs multiple agents on the same project at once, each in its own session.
- Permission rules: Each action resolves to allow, ask or deny, and an --auto flag approves everything not explicitly denied.
- Shareable sessions: Generates a link to any session so a teammate can read exactly what the agent did.
- Terminal, desktop and IDE: One agent available as a terminal interface, a desktop app and an IDE extension.
Strengths
- Fully open source under the MIT licence, so you can read exactly what runs with shell access on your machine.
- Not locked to one vendor: switch models per task, or reuse a Copilot or ChatGPT subscription you already pay for.
- Explicit deny rules still apply even in --auto mode, which gives a hard stop for paths and commands you never want touched.
Weaknesses
- Terminal first, so setup and config take more learning than a one-click IDE assistant.
- The agent runs with your user privileges, and a January 2026 local server flaw (CVE-2026-22812, CVSS 8.8) shows the cost of that: it was fixed in version 1.0.216, so older installs should be updated.
- Two product lines are live (1.x and 2.x), each with its own install route, which makes team-wide pinning more work.
Frequently Asked Questions
What does OpenCode actually cost?
Running the software costs nothing, and some free models are included, so a bring-your-own-key setup costs only what your provider bills. Zen is pay as you go with a $20 starting balance plus a card processing fee, topping up by $20 when the balance drops to $5, and the vendor says it adds no markup. Go costs $10 per month and Go Plus costs $40 per month, each with per-model request limits every 5 hours and monthly usage caps. Enterprise is priced per seat on request.
Is OpenCode safe to use?
OpenCode says it does not store your code or context data, and processing happens locally or through direct calls to your chosen provider. The agent runs shell commands with your privileges, so set permission rules to ask or deny for risky actions, and update past version 1.0.216, which fixed CVE-2026-22812, a local server flaw rated CVSS 8.8. Press coverage dated 2026-10-17 describes an email to Console users about stolen API keys and unauthorized charges, but we found no statement on OpenCode's own changelog or blog when checked on 2026-10-18, so treat it as unconfirmed and rotate any key you gave the Console.
What are OpenCode's closest competitors?
Claude Code suits teams committed to Anthropic's models. Cline is the pick if you want the agent inside VS Code rather than a terminal. Gemini CLI fits people on Google's models, and OpenAI Codex fits people on OpenAI's.
How does OpenCode compare to Claude Code in 2026?
OpenCode is MIT licensed and works across 75+ providers, so you can swap models without changing tools. Claude Code is Anthropic's own agent, built around its own models, which tends to mean a tighter fit but no provider choice. If model flexibility or auditing the source matters most, take OpenCode; if you want the vendor-built path, take Claude Code.
How do you set up OpenCode?
Install it with the vendor's curl script, npm, Bun, Homebrew, Scoop or Docker, then run the /connect command in the terminal interface to add a provider key or sign in. If you are new to model providers, the docs suggest starting with Zen. Open a project folder, describe a task, and approve or deny the first actions it proposes.
Top Alternatives
- Claude Code: Pick OpenCode if you want one open source agent across many model providers; pick Claude Code for Anthropic's own agent loop and models.
- Cline: Pick OpenCode if you work in a terminal or desktop app; pick Cline if you want the agent inside VS Code.
- Gemini CLI: Pick OpenCode if you want to switch between providers; pick Gemini CLI if you plan to stay on Google's models.
- OpenAI Codex: Pick OpenCode for provider choice and an MIT licence; pick OpenAI Codex if you want OpenAI's models in a first-party agent.
- Cursor: Pick OpenCode if you want an open source agent you can point at any model; pick Cursor if you want a polished all-in-one editor.
- Windsurf: Pick OpenCode for terminal work and provider choice; pick Windsurf if you want an editor with its own agent built in.