Open Source Is Back. Goodbye Claude.
BBetter Stack
Computing/SoftwareSmall Business/StartupsInternet Technology
Transcript
00:00:00OpenCode is an open source coding agent that just crossed 180,000 stars on GitHub
00:00:04and it's a direct replacement for tools like ClaudeCode without the glitchy UI and expensive
00:00:10subscriptions. You'll get access to hundreds of models including free and local models which is
00:00:15pretty cool so you're not vendor locked into one platform. Today we're going to pull apart OpenCode
00:00:20and we're going to compare it to some of the paid alternatives and test whether free and open source
00:00:24really holds up when you're trying to push to production. And we cover a huge amount of dev
00:00:29content on this channel so if you want to stay up to date then subscribe.
00:00:37OpenCode is a coding agent that lives in your terminal and it comes with a huge array of tools
00:00:42like bash, edit, grep, webfetch and websearch. These are all the things that made ClaudeCode so useful
00:00:49so seeing them in a free and open source tool is really nice. Once installed we can run the open
00:00:54code command to enter our terminal. Here I'm using ghosty which is one of OpenCode's recommended
00:01:00terminals. You can see immediately the terminal UI here is really clean but the big difference is
00:01:06bringing your own models. By default I'm using a free model called Big Pickle but using the slash
00:01:12models command I can easily switch to other free alternatives or paid models from companies like
00:01:18Google, Anthropic and OpenAI. Hitting control A I can also see a huge list of models which come from
00:01:24models.dev, a directory officially supported by OpenCode. Here you'll have access to models like
00:01:30GLM 5.2 and DeepSeq 4. Even with paid models this can be considerably cheaper than sticking to API
00:01:37only pricing with labs like Anthropic because models like GLM 5.2 are much much cheaper with comparable
00:01:44results to models like Fable 5. GLM 5.2 even beat Fable 5 at web design according to Design Arena. At the
00:01:51time of recording there is no auto mode for OpenCode to select a suitable model based on its assigned task
00:01:57features you'd see in harnesses like Copilot but what you can do instead is configure custom sub
00:02:02agents to use a specific model then call on those for specific tasks and this is a great way to segue
00:02:09into configuring OpenCode's agents. In OpenCode agents come in two types primary and sub-agents.
00:02:15Primary are the top level agents you'd always use you get build and plan exactly as you would expect from
00:02:21harnesses like Copilot or Claude code. You can use tab to change this or just use the slash agents
00:02:27command. So plan is exactly as you would expect for planning large features without actually making those
00:02:32changes and then you can switch to the build agent to actually implement that plan or if you just want
00:02:37to make changes directly stick into the build agent. Sub-agents on the other hand are for specialist
00:02:42tasks and they'll be invoked automatically depending on your prompt but you can tag them manually by using
00:02:48the @ character. By default you'll have access to the general agent for researching complex questions
00:02:53and executing multi-step tasks and then we have the explore agent which has read-only access for exploring
00:03:00large code bases. Now the really interesting and powerful part of OpenCode is that you can configure
00:03:06your own custom agents. So there are a few ways that you can set up custom agents with OpenCode. You can
00:03:12use the JSON file which I just explained earlier and if you want all your config in one place this is the
00:03:17right way to do it. So I'll just enlarge this. We can see that under the agents config we've got a build
00:03:23agent, we've got the plan agent. So these are the ones that come as default but you can change these if
00:03:28you like and then we've got a custom agent here called code review and this just happens to be using Claude
00:03:33Sonic 4 and then we've got a little description of what it does and then a prompt that would
00:03:38automatically be injected into the agents when you ask to perform a specific task. If you don't want
00:03:45all of your config in one place instead you can have a dot OpenCode folder with an agents folder and inside
00:03:51that you can describe the specific agent that you want as a markdown. So here we have the uiengineer.md
00:03:58and the description is it handles UI component implementation styling and front-end engineering tasks.
00:04:04The mode for this would be subagent. We're using GLM 5.2 just because that is very very good at doing
00:04:10design based on the design benchmarks. You can also change the temperature as well to modify the sort of
00:04:15randomness of the output that you would get from the model and then you can also modify any of the
00:04:19default permissions. So in this case we always want to allow edits and always allow bash commands.
00:04:24Within the markdown file you then can give further instructions on exactly what this agent does.
00:04:30So here we say you're a ui engineer specializing in front-end development, you excel at and then you
00:04:34can give as much detail as you like to guide your agent in the correct direction. So next time that
00:04:40you trigger a prompt that requires design work it's going to do exactly what you want. So here as another
00:04:46example we have a review.markdown and this will just do code review for us. In this case we set
00:04:51the temperature as a really low number because we want more of a deterministic output from the model.
00:04:56We provide the same sort of permissions this time we're using Claude Sonic 4 and then we give a little
00:05:01description of what it does review code for quality and best practices. Again the mode is going to be
00:05:07subagent and then again a short or long description to explain exactly what the model does and the more
00:05:12detail you give here the less likely it is to do the wrong thing or hallucinate. Now when i'm working
00:05:17on a ui task i can tag my agent and let glm handle all of the design work. Now here are a few things to
00:05:24improve quality of life when using open code. Flip on lsp in your config and it will auto start the correct
00:05:29language server per file. So the model gets real type information and errors not just raw text and this
00:05:35is off by default so you do need to flip it on. Using the at character also gives you fuzzy file search to
00:05:41pulling files into context. You can use undo or redo to reverse the agents changes and they do stack.
00:05:47Share can give you a link to the whole session for debugging or handing to a teammate. You can also use
00:05:52init which writes an agents.md a checked in file that teaches the agent your projects conventions and this
00:05:58is just exactly like the Claude md file in Claude code. So my verdict here is that the ui is considerably better than
00:06:04Claude code plus it has access to all of the same tools with the added benefit that you can use your
00:06:10own models and configure custom tools and agents. On cost it's actually not black and white if you're on
00:06:15api only pricing then open code wins easily because you have access to so many more much cheaper models
00:06:22but if you're on a subscription with anthropics pro or max plan for example you cannot use third-party
00:06:28harnesses like open code because anthropics specifically block access to them but if you
00:06:32are on that subscription then Claude code bundles in huge discounts compared to raw api rates so if
00:06:38you're on api only pricing then stick to open code but if you do have that subscription then Claude code
00:06:44is the winner on pure cost. If you enjoy hearing about the latest ai tooling then we've done a breakdown of
00:06:50glm 5.2 you can watch here otherwise i've been warren from better stack thanks for watching and i'll see you in the next one