我有点受够 Claude Code 了!
用 OpenCode 直接就开搞!
AI 编程这事儿吧,想入坑其实不难,网上能免费下载的编程智能体多的是。
难的是装完之后——模型呢?
就想先体验体验,又不想一上来就掏钱买 API,只好挨个去找模型厂商,看看谁家有免费额度送。
然后就开始了这段折腾:
大半天过去了,满头汗,一行代码没写,AI 编程还没开始。
直到我遇到了 OpenCode,装上的第一分钟,我愣了一下。
OpenCode 是一个完全开源的 AI 编程 Agent,由 SST 团队(6 人小团队)于 2024 年 7 月发布,MIT 协议,代码完全公开。
它的核心定位是:终端原生的 AI 编程助手,但不绑定任何模型。
你可以理解为:Claude Code 做的那些事——读代码库、改文件、跑命令、执行测试——OpenCode 都能做,但你可以自己选用什么模型,甚至完全不花钱。
| 16万 GitHub Stars | 900+ 贡献者 | 750万 月活开发者 |
6 个人做出这个规模,本身就值得停下来想一想。
这是我觉得 OpenCode 最厉害的地方,也是和 Claude Code 差距最大的地方:它内置了一批可以直接用的免费模型,你不需要去任何一家厂商注册账号、申请 API Key。
200K 上下文窗口,128K 最大输出
目前限时免费——官方正在收集反馈改进模型中
这些是各家厂商推广期主动提供的免费额度,在 OpenCode 里直接
/models 就能选到。 对数据敏感、或者不想有任何云端依赖的人,这条路最稳。
装的过程比我预期的简单得多:
/init 初始化,OpenCode 会分析项目结构,生成 AGENTS.md/models 选 Big Pickle 或任意标了 Free 的模型,直接开始提需求。用起来有两个模式,按 Tab 键 切换:
Build 模式:读文件、改代码、跑命令,闭环执行,交给它直接干
说实话,OpenCode 不是对所有人都完美,但对这几类人来说,它是降维打击:
不需要每天高强度写代码,免费模型完全够用,Ask/Plan 模式就很顺手
零成本起步,Big Pickle 和国产免费模型可以支撑大多数日常任务
Ollama 本地模式,代码完全不出机器,比任何云端方案都干净
多模型并行、随时切换,再也不用盯着用量焦虑
Claude Code 的深度整合确实更顺滑,这种情况继续用没问题
一张表,直接看:
| 对比维度 | Claude Code | OpenCode |
| 是否开源 | ❌ 闭源 | ✅ MIT 开源 |
| 模型支持 | 仅 Claude | 75+ 提供商 |
| 工具费用 | $20~$200/月 | 完全免费 |
| 本地模型 | ❌ 不支持 | ✅ Ollama |
| 数据隐私 | 云端处理 | 可完全本地 |
| GitHub Stars | ~7万 | ~16万 ⭐ |
还有一个意味深长的故事:2026 年 1 月,Anthropic 封锁了 OpenCode 通过消费者账号调用 Claude 的通道,理由是「法律要求」。OpenCode 没有垮,反而顺势把用户引向了更多免费模型、本地模型的方向——结果反而更自由了。
OpenCode 的逻辑是:开放自由换长期主义。
Big Pickle 的 200K 上下文,处理中等规模项目完全不焦虑。给它一个功能需求,先切 Plan 模式让它出方案,确认没问题再切 Build 模式执行,整个节奏很顺。
更重要的是,我不再盯着用量了。以前用 Claude Code,脑子里始终有一根弦:这个任务费不费 token,要不要拆小一点。现在这根弦彻底断了。
体验上当然有区别——Claude Code 的模型集成和细节打磨,确实比 OpenCode 更精致。但对我这种中等强度的使用来说,这点差距,不值一分钱。
6 个人,16 万 Star,750 万月活,全部开源免费。
OpenCode 这个故事让我想起一件事:最好的工具,不一定是最贵的那个。很多时候,它只是那个「装上就能用、不让你操心别的」的那个。
不过我也一直有个问题没想清楚——
这个问题,可能没有标准答案。但它值得每个 AI 编程工具的用户认真想一想。
I'm honestly getting a bit fed up with Claude Code!
No sign-up, no API key hunting, no model configuration — just start coding with OpenCode!
When it comes to AI coding, getting started isn't actually hard. There are plenty of free coding agents you can download online.
The hard part comes after installation — where are the models?
If you just want to try it out without paying for an API upfront, you have to go around to different model providers and see which ones offer free credits.
And that's when the whole ordeal begins:
→ Find a provider, create an account
→ Wait for email verification, phone verification, real-name authentication
→ Go to the console, create an app, apply for an API key
→ Come back, paste the config, realize the format is wrong, fix it again
→ Test it out, not enough quota, switch to another provider…
Half a day gone, sweating bullets, not a single line of code written, and the AI coding hasn't even started.
That is, until I came across OpenCode. The minute I installed it, I was stunned.
No sign-up, no real-name verification, no API key — open it up and the model is right there, ready to write code.
🛠️ What exactly is OpenCode?
OpenCode is a fully open-source AI coding agent, released in July 2024 by the SST team (a 6-person team). It's MIT-licensed, with completely public code.
Its core positioning is: a terminal-native AI coding assistant that isn't tied to any specific model.
Think of it this way: everything Claude Code does — reading codebases, editing files, running commands, executing tests — OpenCode can do too, but you get to choose which model to use, and you can even use it completely free.
The fact that 6 people built something of this scale is itself worth pausing to think about.
🆓 Models available right after install — where does "free" come from?
This is what I think makes OpenCode truly stand out, and where it differs most from Claude Code: it comes with a set of free models you can use directly, with no need to register an account or apply for an API key at any provider.
🥒 Big Pickle · The one most worth talking about
A "stealth model" developed in-house by the OpenCode team, optimized specifically for coding agents. 200K context window, 128K max output. Currently free for a limited time — the team is actively collecting feedback to improve it.
📦 Free API models from providers (marked as Free)
DeepSeek V4 Flash Free, GLM-4.7, MiMo V2.5, Nemotron 3 Ultra, North Mini Code… These are free quotas actively offered by providers during their promotional periods, and you can simply select them in OpenCode with /models.
🖥️ Local Ollama models · Free forever
Runs entirely on your own machine — code never leaves your local environment, no network, no cost. For those who are sensitive about data or don't want any cloud dependency, this is the most reliable path.
⚠️ Let me be honest: models labeled "Free" aren't free forever — providers can adjust their policies at any time. But while they're free to use now, just use them.
⚡ Three steps to get started — literally just three steps
The installation process was far simpler than I expected:
Step 1 · One-command installation
curl -fsSL https://opencode.ai/install | bash
Works on Mac / Linux / Windows, and you can also use npm / Homebrew / Docker
Step 2 · Go to your project directory and run it
cd /your/project && opencode
Type /init to initialize — OpenCode will analyze the project structure and generate AGENTS.md
Step 3 · Pick a model and get going
Type /models to choose Big Pickle or any model marked Free, then start making requests.
There are two modes you can use, toggle with the Tab key:
Plan mode: analyze only, plan only, don't touch a single line of code — great for code review or "sandbox planning" when you're unsure about the direction. Build mode: read files, modify code, run commands, execute end-to-end — just hand it over and let it do the work.
👥 Who is OpenCode best suited for?
Honestly, OpenCode isn't perfect for everyone, but for these groups, it's a complete game-changer:
✅ People who use AI assistance but don't write code heavily day in and day out — the free models are more than enough, and Ask/Plan modes feel very natural
✅ Students / indie developers / side-project programmers — zero cost to start, and Big Pickle plus domestic free models can handle most daily tasks
✅ Developers who care about data privacy — Ollama local mode keeps code entirely off your machine, cleaner than any cloud solution
✅ Heavy users constantly hitting Claude Code quota limits — run multiple models in parallel, switch anytime, no more staring at usage anxiety
❌ Those who want the absolute pinnacle Claude flagship experience and have money to spare — Claude Code's deep integration is indeed smoother, so no reason to switch in that case
⚔️ What exactly is the difference vs. Claude Code?
One table, right at a glance:
And there's a telling story: in January 2026, Anthropic blocked OpenCode's use of Claude via consumer accounts, citing "legal requirements." OpenCode didn't collapse — instead, it steered users toward more free models and local models — and the result was actually more freedom.
Claude Code's logic is: deep integration in exchange for the ultimate experience. OpenCode's logic is: openness and freedom in exchange for long-term value.
💡 I've used it for a while now — let me be honest
Big Pickle's 200K context means handling medium-sized projects without any anxiety. Give it a feature request, switch to Plan mode first to let it propose a plan, confirm it looks good, then switch to Build mode to execute — the whole flow is very smooth.
More importantly, I've stopped watching my usage. With Claude Code, there was always a nagging thought in the back of my mind: is this task going to burn a lot of tokens? Should I break it into smaller pieces? Now that thought is completely gone.
There are of course differences in experience — Claude Code's model integration and attention to detail are indeed more polished than OpenCode. But for someone like me with medium-intensity usage, that gap isn't worth a single cent.
🌙 Final thoughts
Six people, 160K stars, 7.5 million monthly active users, fully open source and free.
The OpenCode story reminds me of something: the best tool isn't necessarily the most expensive one. More often than not, it's the one that "works right out of the box and lets you stop worrying about everything else."
That said, there's a question I still haven't figured out —
When Big Pickle's free period ends, for all those workflows built on free models — will you stick around and pay, or go looking for the next free option?
There may be no standard answer to that question. But it's worth thinking carefully about for every user of AI coding tools.
— A programmer who moved back from Beijing to Xi'an, from big tech back to everyday life
