Run Claude Code against an OpenRouter, Grok, or Kimi subscription instead of an Anthropic API key
Keep a long coding agent session's prompt cache warm to reduce token costs
Monitor and control several AI backend connections from one local dashboard
| torad-labs/splice | mfori/bit-aes | nacode-studios/kdrant | |
|---|---|---|---|
| Stars | 1 | 1 | 1 |
| Language | Kotlin | Kotlin | Kotlin |
| Last pushed | — | 2020-05-11 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | moderate | moderate | moderate |
| Complexity | 3/5 | 2/5 | 3/5 |
| Audience | developer | developer | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires Java 21+, Node 24, an existing Claude Code install, and an API key for whichever backend provider you connect.
splice is a tool that lets you run Claude Code, Anthropic's coding agent, using a subscription to a different AI provider such as ChatGPT, Grok, or Kimi instead of an Anthropic API key. It is an independent project built in Kotlin and is not affiliated with Anthropic, OpenAI, xAI, Moonshot, or OpenRouter. Anthropic considers routing Claude Code to non-Claude models through a custom gateway unsupported, so this is explicitly a use at your own risk setup. The system works through a small local daemon called spliced that runs only on your own machine, never exposed to the outside network. It sits between Claude Code and whichever backend you choose, translating Anthropic's API format into that backend's own format. Each supported provider gets its own command, so instead of typing claude you type claudex for OpenRouter, claude-grok for Grok, or claude-kimi for Kimi, and the daemon routes the request accordingly. Adding a new backend is done by editing a config file rather than writing code. Two features stand out. The first is called the mirror: when a backend model produces a reasoning summary, splice writes that summary back into the visible conversation so it survives across turns and gets carried through when the conversation is compacted to save space. The second is cache warm compaction, which tries to compact long sessions using the same model and reasoning effort as the session itself, keeping the AI provider's prompt cache active and reducing token costs on long coding sessions. The project also ships a dashboard that shows the status of every running backend, current configuration, and usage warnings, all served locally. Setup requires Java 21 or newer, Node 24, Claude Code itself already installed, Python 3, and standard command line tools like curl and bash. It runs natively on Linux and macOS, and on Windows through WSL2. Installation is handled by a script that checks for missing dependencies and offers to fix them, or an agent can be given a prompt to install it for you. This project fits developers who already use Claude Code and want to try it against other AI subscriptions they already pay for, rather than through Anthropic's own billing.
A local proxy that lets you run Anthropic's Claude Code coding agent using a ChatGPT, Grok, or Kimi subscription instead of a Claude API key.
Mainly Kotlin. The stack also includes Kotlin, Java, Node.js.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.