Connect Claude Code & Claude Desktop to DeepSeek V4 with CC Switch

CC Switch is an open-source, cross-platform desktop manager that keeps every AI coding tool on one configuration. The GitHub repository farion1231/cc-switch (MIT licensed, maintained by Jason Young) had roughly 123k stars and 8.3k forks as of August 1, 2026, with v3.19.0 released on July 30, 2026^[3]. It manages provider configuration for eight tools — Claude Code, Claude Desktop, Codex, Gemini CLI, Grok Build, OpenCode, OpenClaw, and Hermes — so you stop hand-editing settings.json, config.toml, and .env files every time you switch providers.

Under the hood it is a Tauri 2 app: a React front end over a Rust backend, with SQLite (~/.cc-switch/cc-switch.db) as the single source of truth. CC Switch writes configs with a temp-file-plus-rename atomic pattern, so a failed switch never corrupts your Claude Code settings.

Claude Code is the one managed tool that supports hot-switching — changes apply without restarting the terminal. Everything else needs a terminal or CLI restart. Switches write to each tool's own config file, which means the VS Code Claude Code extension and the terminal CLI share the same user-level ~/.claude/settings.json.

CC Switch ships official Linux packages: .deb for Debian and Ubuntu, .rpm for Fedora, RHEL, and openSUSE, and a universal .AppImage that runs on any distro. System requirements are Ubuntu 22.04+, Debian 11+, or Fedora 34+. There is no official Flatpak among the release assets.

• On Arch, paru -S cc-switch-bin pulls the package from the AUR.
• Wayland note: the AppImage forces GDK_BACKEND=x11 (XWayland) by default. If clicks do not register or the window renders black, relaunch with CC_SWITCH_GDK_BACKEND=wayland.
• Under tiling compositors like sway or Hyprland the opposite can happen — if the window stops responding on native Wayland, force it back to x11.

# Debian / Ubuntu
sudo apt install ./CC-Switch_*.deb

# Fedora / RHEL / openSUSE
sudo dnf install ./CC-Switch_*.rpm

# Any distro — universal binary
./CC-Switch-*.AppImage

# Arch Linux (AUR)
paru -S cc-switch-bin

# Wayland fix if clicks do not register or the window is black
CC_SWITCH_GDK_BACKEND=wayland ./CC-Switch-*.AppImage

DeepSeek's official Claude Code integration needs no third-party app: point Claude Code at the Anthropic-compatible base URL and set eight environment variables^[1]. Install Claude Code first with npm install -g @anthropic-ai/claude-code (Node.js 18+), then export the block below in your shell or add it to ~/.bashrc.

The base URL must be https://api.deepseek.com/anthropic — not https://api.deepseek.com. The auth token is your DeepSeek API key from platform.deepseek.com/api_keys, and the variable is ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. Create the key once, as shown in the API setup guide, and reuse it everywhere.

The [1m] suffix selects the 1M-context variant of deepseek-v4-pro for the main tiers. Haiku-tier and sub-agent calls run on the cheaper DeepSeek V4 Flash, which keeps fast, high-frequency work off the expensive model.

• ANTHROPIC_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL: deepseek-v4-pro[1m].
• ANTHROPIC_DEFAULT_HAIKU_MODEL and CLAUDE_CODE_SUBAGENT_MODEL: deepseek-v4-flash.
• CLAUDE_CODE_EFFORT_LEVEL=max keeps thinking effort at maximum.
• DeepSeek's awesome-deepseek-agent config adds one more line, CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1, to cut non-essential traffic.

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API Key>
export ANTHROPIC_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max

The GUI path does the same wiring in one form. In CC Switch, add a provider with these exact values^[4] — small differences in the base URL or auth field are the most common reason connections fail.

The base URL ends in /anthropic, not /v1, and carries no trailing slash. Community testing shows that entering https://api.deepseek.com alone can list models but fails requests, and a trailing slash can produce Connection Refused.

The auth variable is ANTHROPIC_AUTH_TOKEN — the same name the official environment block uses. The DeepSeek preset ships inside CC Switch's 50+ built-in presets and usually arrives with model mapping pre-selected.

Claude Desktop connects to DeepSeek through the app's official developer mode. Launch Claude Desktop without logging in, walk Help → Troubleshooting → Enable Developer Mode, then open Developer → Configure Third-Party Inference. Fill in the base URL, your DeepSeek API key, and a model name.

Claude Desktop only accepts claude-sonnet-*, claude-opus-*, and claude-haiku-* role IDs and rejects any other model name. DeepSeek's backend maps those roles to real V4 models automatically. The silent fallback is the catch: an unknown model name quietly becomes deepseek-v4-flash — the classic "I configured Pro but Pro is running Flash" symptom.

CC Switch solves the desktop whitelist with model mapping mode plus local routing. Desktop requests hit a local gateway first; CC Switch translates role IDs to real model names (Sonnet role to deepseek-v4-pro, Haiku role to deepseek-v4-flash) and forwards them to DeepSeek. The request log shows lines like claude-desktop deepseek-v4-flash claude-haiku-4-5 200 — request_model still carries the role ID, which is normal, because that is the name before mapping.

CC Switch v3.18.0 added an OpenCode Go preset that bridges a Go subscription into Claude, Claude Desktop, Codex, and OpenCode^[5]. You paste a plain API key — no OAuth flow — and CC Switch wires the rest. A dedicated pull request added an "OpenCode Go (DeepSeek V4 Flash)" preset for Claude and Claude Desktop specifically.

OpenCode Go is a low-cost subscription for open coding models: $5 for the first month, then $10/month^[7]. It ships with usage limits — when you exceed them you fall back to free models or top up with Zen balance. Because CC Switch accepts the raw API key, you get a DeepSeek-backed Claude Code session without touching Anthropic's billing.

• Open CC Switch and pick the OpenCode Go preset for Claude or Claude Desktop.
• Paste the API key from your OpenCode Zen console.
• Select the DeepSeek V4 Flash entry (or Pro where available).
• Switch provider as usual — the target app restarts its session against the new gateway.

Two issues account for most "Claude Code + DeepSeek is broken" reports. Know both before you switch and you will skip the support threads entirely.

Pitfall one: the reasoning_content 400 on tool calls. DeepSeek thinking mode requires the reasoning_content from a tool-calling round to be passed back verbatim in every later request; drop it and the API returns HTTP 400 ("The content[].thinking in the thinking mode must be passed back to the API.")^[6]. Claude Code and OpenCode can filter this non-standard field, so long agentic loops 400 after a tool call — this is tracked in CC Switch issue #2331 and OpenCode issue #24901. Mitigations: run CC Switch's local proxy with the Rectifier option enabled (Settings → Advanced), which repairs third-party thinking-block format mismatches, or use the community dsv4-cc-proxy, which strips thinking events on the response side and re-adds structure on the request side.

Pitfall two: the [1m] suffix downgrades you in proxy mode. In direct mode, deepseek-v4-pro[1m] is the official 1M-context model name. But when CC Switch local routing or proxy mode is on, the [1m] is forwarded verbatim as part of the model name, DeepSeek does not recognize it, and the backend silently falls back to deepseek-v4-flash^[8]. Your Pro sessions quietly run Flash.

• In CC Switch proxy or routing mode, write deepseek-v4-pro with no suffix; keep [1m] only for direct connections.
• Mixed-vendor sessions are another 400 trigger: if the first half of a conversation used another provider, the history lacks reasoning_content and the next round fails.
• The Anthropic endpoint ignores cache_control, citations, top_k, and image or document content blocks — do not migrate multimodal or heavy-cache workflows as-is.
• Housekeeping: the legacy aliases deepseek-chat and deepseek-reasoner retired on July 24, 2026, so any config still sending them now errors. Use deepseek-v4-pro or deepseek-v4-flash.