OpenCode 2.0 Guide: Setup, Code Mode & First Prompts

Last updated: October 15, 2026

OpenCode 2.0 is the free, open-source AI coding agent that launched on 13 October 2026 and went straight to the top of Product Hunt’s launch charts. If you’ve been renting an AI coding assistant by the month, this is the weekend to test the free alternative: one install, your own API keys, and any model you like — Claude, OpenAI, Google, or a local model that costs nothing at all.

Quick answer: OpenCode 2.0 is the free, open-source AI coding agent (13 Oct 2026). This guide covers install, the 1.x upgrade, Code Mode and 5 first-session prompts.

What Is OpenCode 2.0?

OpenCode is a terminal-first AI coding agent you run on your own machine — and version 2.0 turns it into a shared platform for every screen you code on. Built by Anomaly (the team behind SST), it brings your own API keys (BYOK) to Claude, OpenAI, Google and 75+ providers, with zero markup: you pay each provider’s published rate directly, nothing extra to the tool itself. Version 2.0, now the main release after a three-month public beta, reworks the internals around a single shared background service so the terminal, desktop app, web client and embedded apps all talk to one agent process.

What Is Actually New in OpenCode 2.0?

The headline change is boring-sounding and genuinely useful: one background service for everything. The service starts automatically, and sessions, configuration, integrations, permissions and tool execution all live in one process. Close a terminal window or switch from desktop to web mid-task, and the agent’s long-running work keeps going — you just reconnect. The opencode pair command lets other hosts reach the same service. Beyond that, the release adds a rebuilt API, an embeddable SDK (@opencode/sdk — the launch note says it is still in development), hot-reloading plugins, background subagents and shell commands, and Code Mode, which lets the agent write JavaScript to chain several MCP tools together. One honest catch: plugins written for OpenCode 1.x do not work in 2.0, so upgraders have rewriting to do.

How Do You Install and Upgrade OpenCode 2.0?

Installing from scratch takes one command; upgrading from 1.x takes a deliberate step.

  1. Fresh install (macOS/Linux): run curl -fsSL https://opencode.ai/install | bash in your terminal. Alternatives: brew install opencode or npm install -g opencode-ai.
  2. Already on 1.x? The 1.19.0 release introduced an explicit /update command and, crucially, opencode upgrade --major to move onto OpenCode 2. Updates no longer install themselves silently — you choose when they land, which is exactly what teams want from agent tooling.
  3. Connect a model: add your own provider API key (Anthropic, OpenAI, Google, Groq, AWS Bedrock, Azure, OpenRouter, or local Ollama). OpenCode charges no markup — the token bill comes from the provider you picked. Prefer zero setup? OpenCode Go is a beta subscription bundle of curated open-weight models: $5 for the first month, then $10/month, with usage caps of $12/5h, $30/week and $60/month. OpenCode Zen is the pay-as-you-go credit route from $20.
  4. Start it: cd into your project and run opencode. The terminal UI is native and keyboard-driven; a VS Code extension exists if you want the agent inside your editor.

Desk-researched, not hands-on tested: we verified the install commands and upgrade path against launch coverage and the product’s own docs references, but did not run them on our machines this week.

What Are the 5 Prompts to Run in Your First Session?

The fastest way to evaluate OpenCode is one real task end-to-end — here is a ready first-session plan.

  1. “Scan this repo and explain the architecture in 5 bullets — entry points, data flow, and where the risky parts are.” A comprehension test every agent should pass in under a minute.
  2. “Write a failing test for feature X first, then make it pass.” Forces test-driven behaviour and exposes how well the agent navigates your codebase.
  3. “Find the three slowest functions in this codebase and explain what each is doing.” A profiling-style read task that shows off MCP and tool-chaining in 2.0’s Code Mode.
  4. “Review this diff as a senior engineer: flag bugs, security issues, and anything you’d block in review.” The review mode in miniature.
  5. “Document this module: what it does, its inputs and outputs, and one gotcha a new contributor should know.” Docs mode, and a genuinely useful artefact.

Keep an eye on Code Mode in prompt 3: the agent writing JavaScript to chain MCP tools is the 2.0 feature the paid tools don’t advertise, and it is the clearest sign of where this platform is headed.

OpenCode 2.0 vs Codex vs Claude Code: Which Should You Use?

Pick on cost model and control, not hype: OpenCode wins on price transparency, the paid tools win on polish.

What mattersOpenCode 2.0Claude Code / Codex
Price modelBring your own keys, zero markup; or $10/mo Go betaSubscriptions per seat (Claude Pro/Max, ChatGPT Plus/Codex tiers)
Model choiceAny: Claude, OpenAI, Google, Groq, local OllamaMostly locked to the vendor’s own models
LicenceApache 2.0 open source, forkable, self-hostableProprietary
ClientsTerminal, desktop, web, embeddable SDKTerminal + IDE integrations (Codex has cloud tasks)
MCP toolingCode Mode: agent-written JS chaining MCP toolsMCP supported, less programmatic
Migration cost1.x plugins must be rewrittenStable, vendor-managed upgrades

For a deeper head-to-head of the paid pair, see our Codex vs Claude Code comparison; for what MCP chaining unlocks beyond coding, see our REA guide to MCP-powered agents.

When Does OpenCode 2.0 Beat the Paid Tools?

Our verdict after the launch coverage: switch when you care about the bill, the models, or the code — stay put when you care about convenience. OpenCode wins when you want Claude on Monday and a cheap local model on Tuesday with no reconfiguration; when token bills make subscriptions look expensive (no markup means the maths favours BYOK fast); and when you want the tool itself forkable — Apache 2.0 means the project survives even if the company pivots. The paid tools keep their edge for teams that want zero maintenance, vendor-managed updates, and support contracts. Honest middle path: many developers already run both — paid for the daily driver, OpenCode for the side project where every rupee of API spend is visible.

What Should You Watch Out For?

Three caveats before you commit your weekend. First, the 1.x-to-2.0 plugin break is real — budget an hour to rewrite anything custom. Second, the SDK is explicitly still in development, so don’t build a product on @opencode/sdk yet. Third, community figures floating around (167k+ GitHub stars, millions of monthly users) come from third-party trackers, not Anomaly — impressive, but treat them as directional, not audited. And the evergreen coding-agent reminder: sandbox what the agent can touch. GitHub just made local sandboxing generally available in Copilot for a reason; whatever agent you run, scope its file and network permissions before letting it loose on a work repo. Our Codex security explainer covers the same ground for the OpenAI side.

Frequently Asked Questions

Is OpenCode 2.0 really free?

The software is free and open source under Apache 2.0. You still pay for the AI models you use — via your own API keys at the provider’s rate, local Ollama models at zero cost, or the optional OpenCode Go ($5 first month, then $10/month) and Zen credit bundles.

How do I upgrade from OpenCode 1.x to 2.0?

Update to v1.19.0 first, then run opencode upgrade --major. Expect to rewrite any 1.x plugins — they do not carry over.

What is Code Mode in OpenCode 2.0?

A 2.0 feature that lets the agent write JavaScript to chain several MCP tools in one step, turning the agent from a chat-with-tools into something closer to a programmable platform.

Can I switch between models mid-project?

Yes — that is the core pitch. OpenCode is model-agnostic: Claude, OpenAI, Google, Groq, Bedrock, OpenRouter and local models all work, and you can change models without losing context.

Does OpenCode 2.0 work on desktop and web too?

Yes. The new shared background service is used by the terminal, the web interface and the desktop app, and opencode pair lets other hosts reach the same service — sessions survive switching clients.

Sources and Methodology

Launch facts (13 October 2026 release, background service, opencode pair, @opencode/sdk, Code Mode, background subagents) from AlexTech.ai’s synthesis of The New Stack’s launch coverage; launch date and Product Hunt listing from Product Hunt; upgrade path (/update, opencode upgrade --major, 1.x plugin break) from Oday Bakkour’s coding-agent changelog roundup; install commands, BYOK provider list, OpenCode Go/Zen pricing and community star/user figures from a community-maintained free-AI-tools list on GitHub (third-party figures, treated as directional). Claims rest on launch-week coverage verified against two or more of these sources; community statistics are labelled as third-party and unverified; nothing here is hands-on tested. Verification date: 14 October 2026.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!