MadAppGang/claudish

🔮 Claudish – a universal front‑end for Claude Code

What it is – Claudish is a command‑line tool that lets you run Claude Code (Anthropic’s coding assistant) with any LLM you already have access to. It does this by starting a local server that mimics the Anthropic API, then forwards the request to the provider you specify (OpenRouter, Gemini, OpenAI, Ollama, etc.). In short, you keep Claude Code’s UI and workflow but can swap the underlying model at will.

---

Key capabilities (as described in the README)

Capability	Details
Multi‑provider support	Cloud providers (OpenRouter → 580+ models, Google Gemini, OpenAI, MiniMax, Kimi, GLM, OllamaCloud, OpenCode Zen, Vertex AI) and local runtimes (Ollama, LM Studio, vLLM, MLX).
BYOK (Bring‑Your‑Own‑Key)	Use the API keys you already own; no extra subscription required.
Unified routing syntax	provider@model[:concurrency] (e.g. google@gemini-2.0-flash). Bare model names auto‑detect the correct provider.
Offline mode	Run entirely locally with Ollama or other self‑hosted models – your code never leaves the machine.
Claude Code compatibility	1:1 protocol compliance; any flag you would normally pass to Claude Code (--agent, --effort, …) is forwarded unchanged.
Interactive & headless operation	If you run claudish without a prompt it enters an interactive selector; with a prompt it works in a single‑shot “quiet” mode.
JSON output & streaming	--json gives structured results; normal mode streams Claude Code’s response as it arrives.
Cost tracking & monitoring	Optional --monitor logs traffic to the real Anthropic endpoint; --cost‑track aggregates token usage.
Sub‑agent / AI‑agent support	Built‑in guide (--help‑ai) and file‑based patterns to let Claude Code delegate work to other models without polluting the conversation context.
Cross‑platform	Works with Node.js or Bun; can be run via npx/bunx without installing globally.
Skill integration	claudish --init installs a Claude Code skill (.claude/skills/claudish‑usage/) that automatically applies best‑practice prompts and sub‑agent delegation.

---

Quick start (from the README)

1. Install – one of:

   curl -fsSL https://raw.githubusercontent.com/MadAppGang/claudish/main/install.sh | bash   # Linux/macOS script
   brew tap MadAppGang/tap && brew install claudish                                 # Homebrew
   npm install -g claudish                                                          # npm
   bun install -g claudish                                                          # Bun
   

2. Provide an API key (e.g. export OPENROUTER_API_KEY=sk‑or‑…).
3. Run a command – pick a model with the unified syntax:

   claudish --model oai@gpt-5.3 "implement user authentication"
   # or fully offline
   claudish --model ollama@qwen3-coder:latest "write a sorting function"
   

4. Optional – login without a key – if you already have a paid ChatGPT, Anthropic, Gemini, etc., you can run claudish login and select the subscription; no API key is needed.
5. For agents – use claudish --help‑ai to see the recommended file‑based pattern for delegating work from within Claude Code.

---

Configuration & environment variables

Variable	Provider / purpose
OPENROUTER_API_KEY	OpenRouter (default backend)
GEMINI_API_KEY	Google Gemini
OPENAI_API_KEY	OpenAI
MINIMAX_API_KEY …	MiniMax, Kimi, GLM, Z.AI, etc. (full list in README)
CLAUDISH_MODEL	Default model if you omit --model
CLAUDISH_PORT	Port for the local proxy (random 3000‑9000 if unset)
CLAUDISH_LOCAL_MAX_PARALLEL	Max concurrent local model requests (default 1)
CLAUDISH_CLASSIFIER_PROVIDER	Send Claude Code’s permission‑classifier to a native Anthropic model
…	See the “Settings Reference” section of the repo for the complete list.

---

Who might use it?

• Developers who love Claude Code’s UI but have cheaper or higher‑capacity models elsewhere.
• Teams that want a single CLI that works with both cloud APIs and on‑premise models.
• Tool builders needing JSON‑structured output for automation pipelines.
• Privacy‑conscious users who run everything locally via Ollama or LM Studio.

---

Limitations & notes (from the README)

• The bundled catalog of 25 extra vendors only becomes active when the corresponding API‑key environment variable is present.
• “Monitor mode” proxies to the real Anthropic endpoint for debugging; it does not replace the chosen provider.
• Vision support for non‑vision models is simulated by first sending the image to Claude for a description, then passing that text to the target model.

---

Bottom line

Claudish is a genuine open‑source utility that extends Claude Code’s coding assistant to any LLM you can access, offering a unified CLI, BYOK cost control, offline operation, and deep integration with Claude Code’s skill system. All of the above is taken directly from the project's README.