TOKENICODE DeepSeek Alpha: a desktop Claude Code GUI built around DeepSeek and CC Switch

The mismatch this fork exists to fix

Upstream TOKENICODE keeps Claude naming in the interface. If you point it at a DeepSeek endpoint through CC Switch, the window still says Opus or Sonnet while the request goes somewhere else. The v1.0.7 changelog describes the fix directly: model names are now shown according to the actual provider and model mapping, and when no provider is configured the Claude name is kept rather than being forced to DeepSeek. That is a narrow problem, and the fork is honest about being narrow. It is for people who already use Claude Code as their agent loop but route it through DeepSeek or a third-party gateway, and who got tired of guessing which model a session was really talking to. If you use Claude Code with Anthropic's own API and nothing else, the premise of this fork does not apply to you.

Provider mapping, model resolution and the context snapshot

The architecture is Tauri 2 on the Rust side with a React 19 front end, and the agent work is delegated to a Claude Code CLI subprocess that the app launches. Provider configuration sits between the two: you declare an Anthropic-format or OpenAI-compatible provider with a base URL, API key, model mapping and optional proxy, and the app resolves the outgoing model name from the currently active provider. The v0.10.10-alpha.1 notes describe the failure this prevents: after switching providers the old selection could survive and a local name like gemma4:12b would be sent to DeepSeek or CC Switch. Resolution now follows the active provider, and the bottom model dropdown moves to a model that provider actually offers.

Context accounting is the other half. The app keeps a persisted context snapshot per session, restores it when you reopen a conversation, and de-duplicates repeated message records, which v1.0.8 lists as the fix for context resetting to zero or double-counting. The context gauge includes cached input tokens. A separate setting chooses between a standard 200K context window and a declared 1M window; the 1M option raises the automatic compact threshold from 160K to 800K and passes CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000 to the CLI subprocess so it does not compact early on a 200K assumption. The threshold is also editable by hand in K tokens, with 160K, 400K, 800K and 950K shortcuts.

Installing from a release and configuring a first provider

The README does not document building from source as the supported path. It points at GitHub Releases, where the v1.0.8 assets are a Windows x64 portable executable, a Windows x64 installer, a zip, latest.json and a SHA256 checksum file. The README names the portable build as tokenicode-deepseek-alpha-v1.0.8-windows-x64.exe, so you download the asset for your machine and compare it against the published checksum before running it. The portable executable needs no install; the installer variant registers the app normally.

Once the app is open, the Provider page is where a DeepSeek or CC Switch endpoint gets declared. The README states that the v0.10.12-alpha.1 release settled how the DeepSeek base URL is joined: writing the official form https://api.deepseek.com requests /chat/completions and no longer has /v1 appended for you, while an explicit https://api.deepseek.com/v1 still requests /v1/chat/completions. So the base URL you type is the base URL that gets used, which is worth knowing before you debug a 404.

If you want a local model instead of a hosted one, the Local Model settings page detects a running Ollama service, lists installed models, and can pull a new one by name, for example qwen2.5-coder:7b or deepseek-r1:7b, showing Ollama's progress output. An installed model can be set as a local API provider, which uses the OpenAI-compatible endpoint http://localhost:11434/v1.

Where the fork is thin: platform coverage and release cadence

The download section lists only Windows x64 artifacts. There is no macOS or Linux asset described in the README, and the .env.example file carries Apple signing variables for a macOS local build, which tells you a Mac build is possible from source but is not what the project ships. If your team is on Macs, this fork asks you to build it yourself, and the README does not walk through that.

The release history is short and dense. Three tagged versions landed between 2026-07-30 and 2026-08-04, and the last push to the repository was on 2026-08-04. Several of those entries are fixes to the previous entry: a v1.0.6 avatar and context fix carried into v1.0.7, then v1.0.8 fixing an AskUserQuestion answer that displayed as responded but was never sent to the model. That is a normal shape for a young fork, but it means you should read the changelog for the version you install rather than assuming the newest build has settled.

There is also a version discontinuity worth flagging. v1.0.1 re-published the tree with v0.10.12-alpha.1 as the functional baseline, and the changelog says the features from 0.10.13 through 0.10.17 are not included. If you are comparing against upstream TOKENICODE, the feature sets are not a superset relationship in either direction.

Skills, translation and preview compared with plain Claude Code

The natural alternative is Claude Code in a terminal plus the upstream TOKENICODE GUI. Plain Claude Code gives you the same agent, the same permission modes and the same skills mechanism, with no desktop layer to keep current; what you give up is the Skills panel, the session history with archive and pin, file browsing, the CodeMirror editor and the preview pane. Upstream TOKENICODE gives you those, but its model display stays on Claude naming and it has no complete Skills management panel, per the comparison table in this README.

This fork's Skills panel scans ~/.claude/skills and the project's .claude/skills by default, and v1.0.7 notes that Codex and Agent-specific directories are no longer mixed in, though you can add compatible directories by hand. Skills are invoked with a slash and several can be combined in one call, and the panel can view, edit, enable, disable, copy and locate skill files. Translation is separate and deliberately non-destructive: skill names and descriptions, plus the body of a SKILL.md preview, can be sent to an independent translation API and rendered in Chinese, with results cached locally and the original file untouched. That design choice is the right one. Translating a skill file in place would break the skill for anyone else reading the same directory.

Preview behaviour is a compromise the changelog admits to. Localhost and local files open in the side Preview panel; external sites open in a separate in-app browser window because iframes block them. The README also states that a screenshot button without real capture capability was removed in v1.0.7.

Maintenance, licensing and what an upgrade costs you

The repository is not archived, and the last push was on 2026-08-04, which is recent enough that the project is still moving. What that means in practice is that upgrades are not free. The app has its own Tauri updater signing configuration and switched its in-app update check to this repository, and each release publishes latest.json alongside the binaries, so the update path is self-contained. But the changelog shows cross-cutting changes: v0.10.7 added the context window setting, v0.10.8 made the compact threshold user-editable, v1.0.7 moved Windows key encryption for provider and translation API keys to DPAPI bound to the current user and migrated old provider keys and local translation config. That last one is the kind of change you want to apply deliberately rather than in the background.

On licensing, the project is Apache-2.0 and the README states that the upstream Apache License 2.0 grant and attribution are retained, with LICENSE and NOTICE files in the repository. A fork that keeps the NOTICE file is doing what the licence expects. If you redistribute a modified build, the obligations in Apache-2.0 apply to you as they would to any derivative; that is a question for your own counsel, not something this article can settle.

Editorial conclusion

Adopt it if you run Claude Code against DeepSeek or a CC Switch gateway on Windows and want the model name in the UI to match the model actually being called. Do not adopt it if you need macOS or Linux builds, or if you want a long-stable release line rather than a fork that shipped three tagged versions in the six days before 2026-08-04. Verify two things first: that your provider speaks the Anthropic or OpenAI-compatible format the Provider page expects, and that the release asset you download matches the SHA256 file published alongside it.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. mistydew/tokenicode-deepseek-alpha on GitHub
4. README
5. Releases