Sign in to continue learning
AI Coding Workflows: From Cloud to Local
Intermediate
1h32m
Topics
AI Coding
AI in Software Development
On Device AI
So far you've been working in Claude Code, but choosing Claude Code as a coding agent means you're locked into the Claude ecosystem. And as you move toward local AI, you'll need more flexibility than that. So in this lesson, you'll start taking back control, starting with the coding agent itself. You saw in the last lesson that Claude Code can run Sub-Agents on different models. But these models are hosted and they are from one vendor. Claude Code wasn't built to go far from this approach. Fortunately, there are many coding agents to choose from, including open source options. For this lesson, we'll use OpenCode. It is a popular, powerful, and mature choice. OpenCode also comes in several flavors: a command line version, a desktop app, a web version, and IDE integrations. In this lesson, we'll use the IDE version through ACP, which is just the connection which lets the coding agent work directly inside your editor. But you should try the different options and find the one that fits your workflow best. You'll click New Chat to start a new chat window. It defaults to the Claude Agent. That menu though, lets us install from the ACP Registry. There you'll find a long list of agents. Scroll down to OpenCode and click Install. Then click Apply and click OK to close those settings. Back in the chat window, OpenCode is now the agent. You'll then see a listing of models that OpenCode themselves host. Let's make sure to choose Big Pickle, one of OpenCode's currently free hosted demo models. It's a quick, low-stakes way to test that OpenCode is installed and working. You'll enter a simple prompt just to confirm OpenCode is set up. The engine changed from Claude, but the AI chat should be similar. Changing the coding agent was the first layer. But the agent still needs a model to talk to. Now that you've moved away from Claude Code, you're no longer restricted to Anthropic models. OpenCode can connect to many different models, but setting each one up separately means juggling accounts, keys, and billing. That's where OpenRouter comes in. It gives OpenCode one API connection to a catalog of models so you can choose the model while OpenRouter handles the rest. OpenRouter charges based on token usage, so later you'll add a small amount of credits for this course. Now, the models are still running in the cloud, but You now control which model your coding agent uses. First, go to the OpenRouter site and create a login. Once you've created the account, this is what the dashboard will look like. Once logged in, you have a two-step process. You first add some credits to your account. For this course, you can add the equivalent of $10 and that should be plenty. I already have an account with credits. For your account, click Add Credits and follow the screen to provide $10. Next, you need to make an API key. This is how you will connect your OpenCode into the use of OpenRouter. Finally, copy this API key to the clipboard. It's important to do this now as you can't come back later and get it. Also, API keys are a secret. Don't share this with anyone. Back in the IDE, you can configure OpenCode to add OpenRouter as a provider. You can choose to edit the obscure OpenCode JSON file yourself, but hey, you have OpenCode. Just tell it to set up OpenRouter. Once done, you start a new OpenCode chat. After a couple of seconds, voila, you have the OpenRouter models. What does OpenRouter give you? Access to lots of models hosted at inference providers, all through a single account and API key. You now have OpenCode connected to OpenRouter as both a router and inference provider. Now select the model you want to use. Deepseek V4 Pro in particular is a good fit for what you're doing. At the time of this recording, compared to Anthropic models, Deepseek Pro is cheaper than Haiku, but more powerful than Sonnet. OpenRouter gives you access to Deepseek V4 Pro and Flash through the credits you just added. You'll start with Pro and now let's get OpenCode to use it. First, you change the current chat to use DeepSeek Pro. Click the model picker, find DeepSeek V4 Pro in the long list and select it. You can now do a prompt confirming OpenCode works with OpenRouter using our target model. That long model list is cumbersome. Let's do a bonus track. You'll use a prompt to configure OpenCode to only show what you want. You'll be using DeepSeek Flash, so you include that as well. OpenCode will then edit its configuration file. Your prompt also defaults the model reasoning level to medium. Let's confirm this works. Start a new chat and the model picker should default to your last used model, Deepseek Pro, and the list should only include two models at the default reasoning. Send a hello prompt and confirm we're using one of our two models. You're now ready to return to development. Looking back again at your matrix, this lesson changed quite a bit, agent, provider, and model. This provides control, but this choice shifts responsibility to you. Let's face it, the choice can be overwhelming. So many good agents to choose from. So many models, each with variations in size. And you haven't even started yet on running these things semi-locally or locally. As you split up the monolith, you'll find multiple places to have different model choices for parts of the workflow. The hope in this course isn't to say, do this. It's to teach you how to think about these choices. Again, everything is changing quickly. Changing from Claude Code to another agent is a big step. But so far, it's just a setup. In the next lesson, you'll put it to work, building the same project again, this time with your new agent and model.
