Step-by-Step Guide: Connecting DeepSeek to Claude Code Desktop

Why Connect Claude Code Desktop to Domestic AI Models

Claude Code, Anthropic's AI coding tool, has become a go-to utility for many developers thanks to its strong code comprehension and generation capabilities. Anthropic was founded in 2021 by Dario Amodei, former VP of Research at OpenAI, with a focus on AI safety research. Claude Code supports the full programming workflow — code generation, debugging, and refactoring — and goes beyond simple conversation by using Tool Use to perform real actions like file I/O and terminal execution, effectively evolving into an intelligent agent. For users in China, however, two pain points are hard to ignore: official Claude models are expensive, making frequent use costly over time, and account registration and network access barriers put off many newcomers.

This article provides a hands-on walkthrough centered on Claude Code Desktop, with the core goal of replacing the official model with DeepSeek or other domestic alternatives — preserving Claude Code's tooling capabilities while dramatically cutting costs. The process covers five key steps: installing the desktop app, bypassing account login, applying a Chinese localization patch, connecting DeepSeek, and loading custom Skills. Follow along and you should be up and running in about 10 minutes.

A quick heads-up: the third-party switching tools and localization patches described here are community solutions. Routing Claude Code through a different model is an unofficial modification. Pay attention to version compatibility and data security when using these approaches.

Step 1: Install the Desktop App and Enable No-Account Mode

Start by visiting the official Claude Code page and finding the "Get Claude Code" entry point. Hovering over it reveals several usage options; the Desktop version is listed first and is what we need. The page auto-detects your OS and serves the appropriate installer — Windows users will get a Windows build directly.

Once downloaded, run the installer. It's a straightforward process with no extra configuration needed. After installation, you'll land on a login screen where Google sign-in and other methods are available — but our approach is to skip the account entirely.

The key to no-account usage is enabling Developer Mode: click the three-line hamburger menu in the top-left corner, go to Help, scroll down to Enable Developer Mode, and click to turn it on. A confirmation prompt will appear; confirm and Claude Code will restart automatically. After the restart, the door is open for connecting third-party models.

Step 2: Switch to a Domestic Model Using CC Switch

With no-account mode active, you need one critical middleware tool — CC Switch. It's a piece of software for managing AI tool workflows. Its core mechanism is launching a local proxy service that intercepts API requests from Claude Code and forwards them to third-party model endpoints like DeepSeek. This pattern is technically known as an "API compatibility layer" or "model gateway" — as long as the target model implements an interface compatible with the Anthropic API (typically in OpenAI-compatible format), the client can't tell the difference. In short, CC Switch lets you seamlessly swap between models.

One critical version note: you must use CC Switch version 3.15.0 or higher. Only newer versions include built-in support for Claude Code Desktop. If you previously installed an older version or downloaded a package below this version, update it first — otherwise the desktop integration won't work.

Once installed, open CC Switch. At the top of the interface, alongside the existing CLI entry point, you'll now see a Desktop option marked with a small computer icon — that's your indicator of desktop support. Here's how to configure it:

1. Click Settings on the left, go to Routing, find Local Routing, and click the down arrow to enable it.
2. Go back and select Desktop, click Add on the right, and choose the model to load (this guide uses DeepSeek as the example).
3. Enter your DeepSeek API Key.

Getting an API Key is straightforward: go to the DeepSeek Open Platform, navigate to the API Key management page in the documentation, create a new key specifically for Claude Code Desktop, and paste it in.

About DeepSeek: DeepSeek is a family of large language models developed by DeepSeek (Hangzhou) Artificial Intelligence Co., Ltd. Its API pricing is significantly lower than international counterparts like GPT-4 or Claude, while benchmark performance on tasks like code generation and mathematical reasoning approaches or even surpasses some top closed-source models. The DeepSeek Open Platform provides a standard REST API compatible with OpenAI's calling format, making it easy to integrate with any AI client tool that supports custom endpoints. Common versions include DeepSeek-V3 (general conversation and code) and DeepSeek-R1 (enhanced reasoning). Create separate keys for different use cases to simplify management and usage tracking.

Scroll down after entering the key and you'll see an optional context size setting. Enabling it expands the context window to approximately 1M tokens — recommended. The context window determines how much text the model can "see" in a single session: when working with large codebases, the model needs to understand content across multiple files; during refactoring, dependencies must all be present in the same context. Expanding to ~1M tokens means the model can process dozens of mid-sized source files at once, greatly reducing comprehension errors caused by context truncation. Note that a larger context also means more tokens consumed per request, which increases API costs. Once configured, click Add, then be sure to click Start — the configuration only takes effect when CC Switch is running. You can then minimize it to the background.

Step 3: Complete Third-Party Configuration and Verify the Connection

Back in Claude Code, click the top-left menu — you'll now see a Developer option with a Configure Third-Party entry inside. Opening it, you'll find the fields have already been auto-populated by CC Switch. No manual input needed. This is also why you shouldn't open this configuration panel before CC Switch is properly set up and running — it only auto-fills once CC Switch is active.

Click Apply and restart. Claude Code will now launch with DeepSeek as its underlying model. In the chat interface, you'll see selectable model versions including Flash and Pro. Choose Flash, send a test message like "Who are you?", and if you get a normal response, the integration is working. Note that the model may still identify itself as Claude — that's expected behavior. The system prompt and client logic remain unchanged; only the model actually generating responses has been swapped out, so it doesn't affect practical use.

Step 4: Apply the Chinese Localization Patch

By default, Claude Code Desktop is entirely in English, which isn't ideal for users in China. The community has produced a localization patch supporting both Windows and Mac, with options for Simplified Chinese, Traditional Chinese, and Traditional Chinese (Hong Kong).

Download the corresponding archive, extract it, find the .bat batch file inside, and double-click to run it. The script offers five options: install Simplified Chinese, install Traditional Chinese, install Traditional Chinese (Hong Kong), uninstall and restore the original, or exit. Press 1 and hit Enter to install Simplified Chinese.

During the process, the script will automatically close Claude Code (manually closing it beforehand is safer). It restarts automatically once done. On relaunch, UI elements like "New Task", "Projects", "Scheduled Tasks", and "Custom" will all appear in Chinese — localization successful. Previous conversation history is fully preserved.

Step 5: Load Custom Skills in the Desktop Version

Loading Skills in the desktop version differs noticeably from the CLI version. In the CLI, you simply place Skill files in the skills folder under the .claude directory. In the desktop version, Skills must be installed as a compressed archive.

In Claude Code, Skills function like reusable "prompt template packages" or "workflow presets." Technically, a Skill is typically a structured file (commonly Markdown or JSON) containing system instructions, tool configurations, and behavioral constraints. It tells the model what processing logic, output format, and behavioral rules to follow for a given task type. This concept is similar to OpenAI's GPTs or Coze workflow bots — both use predefined instruction sets to customize AI behavior. For developers, creating dedicated Skills for different project types (e.g., frontend development, API design, technical writing) can significantly improve consistency and efficiency on repetitive tasks.

To load a Skill: create a new task, click the plus icon next to the input field, then navigate to Skills → Manage Skills → Add Skill → Upload Skill. The desktop version supports .md files, compressed archives, or .skill files.

Using the demo IvanCreative creative Skill as an example: package it into a zip file and drag it in to upload. Once uploaded, type a forward slash / in the chat box to select and invoke the skill. The model will process and respond according to the Skill's preset logic. Note that if the Skill involves image generation, additional configuration is required — this guide only validates text-based Skills.

Summary and Practical Tips

After completing this entire process, Claude Code Desktop transforms from a tool locked to an official account and expensive models into a low-cost development assistant that connects to DeepSeek and other domestic models, displays a Chinese interface, and supports custom Skills. For developers on a tight budget who still want the Claude Code workflow, this is a genuinely practical path.

A few critical reminders: CC Switch must be version 3.15.0 or higher — this is a prerequisite for desktop integration. After configuration, you must click Start and Apply, then restart for changes to take effect. Close the software manually before applying the localization patch to avoid issues.

Also approach this setup with realistic expectations. Third-party routing and localization patches are not officially supported; monitor for version updates and stability issues over time. When handling sensitive code, carefully evaluate the data flow: requests routed through CC Switch are ultimately sent to the selected model provider's servers. Make sure you understand the relevant data handling policies before use.

Key Takeaways