Claudish does AI differently — it acts as a TypeScript CLI proxy that translates and routes requests from Claude Code to a wide variety of AI models, making Claude’s protocol compatible beyond Anthropic’s native ecosystem. This approach opens up access to over 580 models across multiple providers and local inference engines, all with a consistent interface.
what claudish does and its architecture
At its core, Claudish is a command-line interface tool written in TypeScript that sits between Claude Code, an AI agent orchestration CLI, and a multitude of AI providers. It intercepts API requests from Claude Code and translates them into an Anthropic-compatible format, effectively acting as a universal adapter for AI models.
The key architectural concept is the provider@model routing syntax. This allows users to specify which AI provider and model to target in a uniform way. For example, a request can be routed to openai/gpt-4 or google/gemini-1 or even local models like ollama/stable-vicuna.
Supported providers include OpenRouter (which aggregates 100+ models), direct APIs like Google Gemini, OpenAI, MiniMax, Kimi, GLM, and local inference engines such as Ollama, LM Studio, and vLLM. This design means Claudish can work across cloud and local setups, supporting a BYOK (Bring Your Own Key) model with no extra subscriptions.
A standout feature is the vision proxy capability. When using non-vision models that lack native image understanding, Claudish routes images through Claude itself to generate textual descriptions. This makes it possible for any supported model to interpret visual content indirectly, democratizing multimodal AI workflows.
Under the hood, Claudish integrates with Claude Code CLI, requiring it as a prerequisite. It extends the Claude Code protocol to support multi-provider environments and sub-agent delegation patterns, enabling complex AI agent workflows where different sub-tasks can be routed to specialized models.
technical strengths and tradeoffs
The main strength of Claudish is its extensive support for over 580 AI models through multiple providers and local inference engines, all accessible via a consistent CLI interface. This flexibility means you’re not locked into a single vendor or API, which is valuable in a rapidly evolving AI landscape.
The provider@model routing syntax is a simple yet powerful abstraction that makes switching or combining models straightforward. The codebase is TypeScript, which helps maintain type safety and developer experience across the proxy’s complex translation logic.
The vision proxy mechanism is clever and practical. By generating image descriptions with Claude for models that don’t natively support vision input, Claudish enables multimodal use cases without requiring every provider to have vision capabilities. This tradeoff means some latency and additional steps, but it unlocks a broader range of applications.
The sub-agent delegation pattern supports more advanced AI workflows by allowing different agents or sub-tasks to be delegated to different models. This modularity aligns with trends in AI orchestration but adds complexity in setup and debugging.
A tradeoff to note is the dependency on Claude Code CLI; Claudish cannot operate independently. Also, managing multiple API keys and local model setups can be cumbersome for casual users. While the BYOK approach avoids subscription fees, it requires users to have valid keys and infrastructure.
From a code quality perspective, the repo is well-structured with clear routing logic and modular providers. The use of TypeScript is a plus for maintainability, though the project targets experienced developers who are comfortable managing CLI tools and API integrations.
quick start
installation
### prerequisites
- Claude Code - Claude CLI must be installed
- At least one API key:
- OpenRouter API Key - Access 100+ models (free tier available)
- Google Gemini API Key - For direct Gemini access
- OpenAI API Key - For direct OpenAI access
- OllamaCloud API Key - For cloud-hosted Ollama models (`oc/` prefix)
- Or local models (Ollama, LM Studio) - No API key needed
### other install options
**Use without installing:**
```bash
npx claudish@latest --model x-ai/grok-code-fast-1 "your prompt"
bunx claudish@latest --model x-ai/grok-code-fast-1 "your prompt"
Install from source:
git clone https://github.com/MadAppGang/claudish.git
cd claudish
bun install && bun run build && bun link
quick start
# Initialize Claudish skill (first time only)
claudish --init
# Install dependencies
bun install
# If Claude Code CLI is not installed
npm install -g claude-code
This quick start sequence sets up Claudish with the necessary dependencies and prepares it for use with Claude Code.
verdict
Claudish is a solid tool for developers needing a flexible multi-model AI CLI that integrates deeply with Claude Code. Its support for a vast array of models and providers makes it ideal for experimentation, multi-provider workflows, or bridging local and cloud AI engines.
The vision proxy feature is a neat workaround for enabling image understanding on models that don’t support it natively, though it adds some complexity and potential latency. Dependency on Claude Code CLI and the need to manage multiple API keys/local setups are the main limitations.
If you’re comfortable with CLI tooling, API keys, and want to avoid vendor lock-in while experimenting with multimodal AI workflows, Claudish is worth exploring. It’s not for those seeking simple turnkey AI CLI solutions, but it shines in flexible, real-world AI orchestration scenarios where you want control over model routing and capabilities.
Related Articles
- Inside Google Gemini CLI: a terminal-first AI agent with extensible Model Context Protocol — Google Gemini CLI is a TypeScript-based terminal AI agent offering direct Gemini model access, extensibility via MCP, an
- OpenAI Codex CLI: local-first AI coding assistant with ChatGPT integration — OpenAI Codex CLI brings AI coding assistance local to your terminal, integrating with ChatGPT plans for powerful hybrid
→ GitHub Repo: MadAppGang/claudish ⭐ 846 · TypeScript