qai

Multi-provider launcher for Claude Code with unified configuration management.

Overview

QAI transparently routes Claude Code to different AI providers by intercepting API calls and translating between provider-specific formats. It supports both direct Anthropic-compatible endpoints and NVIDIA NIM endpoints that require protocol translation.

Key Features

• Multi-provider support: Switch between different AI providers via command-line
• Automatic proxy: NVIDIA NIM endpoints automatically go through a translation proxy
• Unified configuration: Single YAML file manages all provider credentials
• Clean process management: Proper cleanup of proxy and Claude processes
• Environment injection: Provider-specific environment variables for Claude Code

Architecture

┌─────────────────┐        ┌─────────────────┐        ┌──────────────┐
│   Claude Code   │───────▶ │      QAI        │───────▶ │   Provider   │
│                 │  STDIO  │                 │  HTTP   │              │
└─────────────────┘         └─────────────────┘         └──────────────┘
                            │  ↑                         ↑
                            │  │ Proxy (NVIDIA only)     │
                            └──┴─────────────────────────┘

Two modes:

• Direct mode: Claude Code connects directly to provider (Anthropic-compatible)
• Proxy mode: QAI starts an HTTP proxy that translates between Claude Code and NVIDIA NIM

Setup

Create configuration file:

mkdir -p ~/.config/qai
touch ~/.config/qai/config.yaml

Configuration Syntax

# Optional: Set default provider
default: zai

# Provider definitions
providers:
  # Provider name (used with -p flag)
  zai:
    # Provider type: "zai" (default, Anthropic-compatible) or "nvidia"
    provider: "zai"
    # API authentication token
    token: "your-token-here"
    # Optional: Environment variables to pass to Claude Code
    env:
      ANTHROPIC_DEFAULT_OPUS_MODEL: "GLM-5"

  nim:
    provider: "nvidia"
    token: "nvapi-xxxxxxxxxxxxxxxxxxxxxxxxxx"
    # Optional: Custom base URL (defaults to NVIDIA endpoint)
    # base_url: "https://custom-nvidia-endpoint/v1"

  # Custom Anthropic-compatible endpoint
  custom:
    provider: "zai"
    token: "sk-ant-xxxxxxxx"
    base_url: "https://api.custom-provider.com/v1"

Provider Types

Type	Description	Needs Proxy	Base URL Default
zai	Anthropic-compatible (Claude API format)	No	https://api.z.ai/api/anthropic
nvidia	NVIDIA NIM (requires translation)	Yes	https://integrate.api.nvidia.com/v1/chat/completions

Usage

Launch Claude Code

# Use default provider
qai claude

# Specify provider by name
qai -p zai claude
qai -p nim claude

# Pass arguments to Claude Code
qai claude --help
qai -p zai claude /path/to/repo

Debug Proxy

Start the translation proxy without launching Claude:

qai -p nim proxy

Output will show the local proxy port (e.g., :58473):

nim proxy on :58473

You can then test the proxy:

curl http://127.0.0.1:58473/v1/messages \
  -H "X-Api-Key: test" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-3-5-sonnet","messages":[]}'

Configuration Details

Required Fields

Field	Required When	Description
token	Always	API authentication token
base_url	Custom endpoints	Override default provider endpoint

Optional Fields

Field	Description
provider	Provider type or "zai" (default)
env	Additional environment variables for Claude Code
model	Model name to pass to Claude Code

Environment Variables

The following environment variables are automatically set:

• ANTHROPIC_AUTH_TOKEN: Set to provider token
• ANTHROPIC_BASE_URL: Set to proxy URL (for NVIDIA) or custom base URL
• ANTHROPIC_MODEL: Set to provider's model if specified

Additional variables from the provider's env field are merged into Claude Code's environment.

Troubleshooting

"No provider" error

Create the configuration file or specify a provider with -p:

qai -p zai claude

"Unknown provider" error

Check that the provider name in your config matches what you're using with -p.

"No token" error

Add a token field to the provider configuration in ~/.config/qai/config.yaml.

Proxy not available

The proxy command only works with providers that require a proxy (currently only nvidia type).

Process cleanup issues

QAI registers signal handlers for SIGINT, SIGTERM, and others to ensure proper cleanup. If processes remain, manually kill them:

# Find proxy processes
ps aux | grep nvidia-claude-proxy

# Kill process tree
kill -9 -<PID>