7.4 KiB

Raw Blame History

Claudish Documentation

Run Claude Code with any AI model. Simple as that.

You've got Claude Code. It's brilliant. But what if you want to use GPT-5 Codex? Or Grok? Or that new model everyone's hyping on Twitter?

That's Claudish. Two ways to use it:

CLI Mode - Replace Claude with any model:

claudish --model x-ai/grok-code-fast-1 "refactor this function"

MCP Server - Use external models as tools inside Claude:

"Claude, ask Grok to review this code"

Both approaches, zero friction.

Why Would You Want This?

Real talk - Claude is excellent. So why bother with alternatives?

Cost optimization. Some models are 10x cheaper for simple tasks. Why burn premium tokens on "add a console.log"?

Capabilities. Gemini 3 Pro has 1M token context. GPT-5 Codex is trained specifically for coding. Different tools, different strengths.

Comparison. Run the same prompt through 3 models, see who nails it. I do this constantly.

Experimentation. New models drop weekly. Try them without leaving your Claude Code workflow.

60-Second Quick Start

Step 1: Get an OpenRouter key (free tier exists)

# Go to https://openrouter.ai/keys
# Copy your key
export OPENROUTER_API_KEY='sk-or-v1-...'

Step 2: Pick your mode

CLI Mode - Replace Claude entirely

# Interactive - pick a model, start coding
npx claudish@latest

# Single-shot - one task and exit
npx claudish@latest --model x-ai/grok-code-fast-1 "fix the bug in auth.ts"

MCP Mode - Use external models as Claude tools

Add to your Claude Code settings (~/.config/claude-code/settings.json):

{
  "mcpServers": {
    "claudish": {
      "command": "npx",
      "args": ["claudish@latest", "--mcp"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

Then just ask Claude:

"Use Grok to review this authentication code"
"Ask GPT-5 Codex to explain this regex"
"Compare what 3 models think about this architecture"

CLI vs MCP: Which to Use?

Scenario	Mode	Why
Full coding session with different model	CLI	Replace Claude entirely
Quick second opinion mid-conversation	MCP	Tool call, stay in Claude
Batch automation/scripts	CLI	Single-shot mode
Multi-model comparison	MCP	`compare_models` tool
Cost-sensitive simple tasks	Either	Pick cheap model

TL;DR: CLI when you want a different brain. MCP when you want Claude + friends.

Documentation

Getting Started

Quick Start - Full setup guide with all the details

Usage Modes

Interactive Mode - The default experience, model selector, persistent sessions
Single-Shot Mode - Run one task, get result, exit. Perfect for scripts
MCP Server Mode - Use external models as tools inside Claude Code
Monitor Mode - Debug by watching real Anthropic API traffic

Models

Choosing Models - Which model for which task? I'll share my picks
Model Mapping - Use different models for Opus/Sonnet/Haiku roles

Advanced

Environment Variables - All configuration options explained
Cost Tracking - Monitor your API spending
Automation - Pipes, scripts, CI/CD integration

AI Integration

For AI Agents - How Claude sub-agents should use Claudish

Help

Troubleshooting - Common issues and how to fix them

The Model Selector

When you run claudish with no arguments, you get this:

╭──────────────────────────────────────────────────────────────────────────────────╮
│  Select an OpenRouter Model                                                      │
├──────────────────────────────────────────────────────────────────────────────────┤
│  #   Model                             Provider   Pricing   Context  Caps       │
├──────────────────────────────────────────────────────────────────────────────────┤
│   1  google/gemini-3-pro-preview       Google     $7.00/1M  1048K    ✓ ✓ ✓      │
│   2  openai/gpt-5.1-codex              OpenAI     $5.63/1M  400K     ✓ ✓ ✓      │
│   3  x-ai/grok-code-fast-1             xAI        $0.85/1M  256K     ✓ ✓ ·      │
│   4  minimax/minimax-m2                MiniMax    $0.60/1M  204K     ✓ ✓ ·      │
│   5  z-ai/glm-4.6                      Z.AI       $1.07/1M  202K     ✓ ✓ ·      │
│   6  qwen/qwen3-vl-235b-a22b-instruct  Qwen       $1.06/1M  131K     ✓ · ✓      │
│   7  Enter custom OpenRouter model ID...                                        │
├──────────────────────────────────────────────────────────────────────────────────┤
│  Caps: ✓/· = Tools, Reasoning, Vision                                           │
╰──────────────────────────────────────────────────────────────────────────────────╯

Pick a number, hit enter, you're coding.

Caps legend:

Tools - Can use Claude Code's file/bash tools
Reasoning - Extended thinking capabilities
Vision - Can analyze images/screenshots

My Personal Model Picks

After months of testing, here's my honest take:

Task	Model	Why
Complex architecture	`google/gemini-3-pro-preview`	1M context, solid reasoning
Fast coding	`x-ai/grok-code-fast-1`	Cheap ($0.85/1M), surprisingly capable
Code review	`openai/gpt-5.1-codex`	Trained specifically for code
Quick fixes	`minimax/minimax-m2`	Cheapest ($0.60/1M), good enough
Vision tasks	`qwen/qwen3-vl-235b-a22b-instruct`	Best vision + code combo

These aren't sponsored opinions. Just what works for me.

Questions?

"Is this official?" Nope. Community project. OpenRouter is a third-party service.

"Will my code be secure?" Same as using OpenRouter directly. Check their privacy policy.

"Can I use my company's private models?" If they're on OpenRouter, yes. Option 7 lets you enter any model ID.

"What if a model fails?" Claudish handles errors gracefully. You'll see what went wrong.

7.4 KiB Raw Blame History