Authored by Anil Kumar
When we think about coding in Salesforce, what typically comes to mind is opening the Developer Console and writing Apex, Visualforce pages, or Lightning Web Components. For more advanced development — including version control and Git integration — the tool of choice for most Salesforce developers has been Visual Studio Code. This has largely been the standard workflow since Salesforce launched, used by customers and developers across the globe.
But recent advances in AI are starting to reshape that picture. Where AI on the platform was once mostly rules-based or predictive — think Einstein’s lead scoring or next-best-action recommendations — it has now shifted toward Generative AI. This new generation of AI can draft notes, summarize calls and cases, generate code, and much more, all powered by large language models (LLMs).
As more AI features were added to the Salesforce platform, a similar evolution happened with Visual Studio Code. We now have the ability to integrate various LLM-powered chats or terminals directly within VS Code. One such advancement is the integration of Claude Code Anthropic’s AI command-line interface, directly inside the VS Code terminal.
Claude Code is an AI-powered command-line interface that runs directly inside your VS code terminal. It is aware of your entire project structure scanning dependencies, reading files, and understanding how everything connects. This concept of full project awareness is the foundation of Claude Code.
Note: To use Claude Code within Visual Studio Code, you need an active Claude subscription
And to start using claude pro in your project you need four things :
For Claude to understand your open project — and what kind of codebase it’s working with — it needs to explore and read through your project files. Keep in mind that this process can consume a significant number of tokens, so it’s worth choosing the right model for the task to manage costs efficiently.So consider using a lighter-weight model like Haiku for routine tasks, and reserve Sonnet or Opus for more complex reasoning
Type the command in claude terminal /init
Running the /init command invokes a built-in subagent called Explore, which by default runs on Haiku ,a lighter, faster model. Its job is simply to scan the codebase and generate a summary, which it sends back to the main agent. Claude then uses that summary to generate a CLAUDE.md file , essentially the project’s memory. This file captures Claude’s understanding of your naming conventions, best practices, custom objects, integrations, schemas, and more.
Because this context loads automatically at the start of every session, you won’t need to keep re-explaining the same project details each time you give Claude instructions in the terminal. In a Salesforce context, this means Claude picks up that it’s working within a Salesforce project and follows that framework accordingly.
A realistic example from one my salesforce project look like :
Claude picks up from this that the project is a Salesforce project with Agentforce capabilities (and more). From there, it knows which CLI commands to run — for example, if you type “deploy this generated file to my default org” in the terminal, it maps that to the right sf command automatically.
You can also add the Agentforce Vibes rules to your CLAUDE.md file so Claude follows them ,Agentforce Vibes rules are Salesforce’s own best-practice guidelines.
During these activities, it’s always a good idea to check your token consumption, especially if you’re using a Claude Pro subscription. You can do this by typing the /usage. And It shows your current usage and when your limit will reset if you run out of tokens.
Now that everything’s set up, we can actually put Claude Code’s terminal to work. That means giving it plain-language prompt instructions — Claude will navigate to the project folder, locate the relevant files or folders, and explore their current state. For example, if you ask it to generate a new custom object with a set of fields, it will create that object and its fields, then show you the newly generated metadata file. This gives you a chance to review the changes before instructing Claude to deploy them to your default org.
So, if I give Claude the following instruction in the terminal:
Create a new custom object called “ClaudeObject “with the fields – ‘AccountName’ — Text field, length 255 and ‘Type’ Picklist field with values: New, Old, Closed (with default as New).
So if you follow the flow in the image above, it goes to your currently open project, reads those files, and creates a new object with the fields I instructed in the terminal. You can see it obey that instruction and create the fields exactly as I specified. It also gives you a chance to review those files right in the terminal, and if you see that it’s exactly what you need, you can then instruct it to deploy to your default org.
Now if I run this to deploy, it will go back to the CLAUDE.md file and check the Salesforce CLI commands it needs to follow for the deployment.
And if we open our default org to check whether this object has been deployed:
Conclusion
What this walkthrough really shows is a shift in where Salesforce development happens. Instead of hand-writing metadata XML or clicking through Setup, you’re describing what you want in plain English and Claude Code handles the object creation, field types, picklist values, and deployment commands, all while staying consistent with the conventions already captured in your CLAUDE.md file. The terminal becomes less of a command interface and more of a collaborator that already knows your org.
This is just the start — I’ll be sharing more hands-on walkthroughs like this as I keep exploring Claude Pro and Salesforce together. If you found this useful, follow and subscribe to the blog