cc-copilot-bridge

TL;DR: Bash script that routes Claude Code CLI through multiple AI providers. Switch between Anthropic Direct API, GitHub Copilot (via copilot-api proxy), or Ollama local with simple aliases (ccd, ccc, cco).

📖 New to Claude Code? Check out the Claude Code Ultimate Guide for comprehensive documentation, tips, and best practices.

StarMapper

🎯 What Is This?

A multi-provider router for Claude Code CLI that lets you switch between AI backends with simple aliases.

Three Providers, One Interface

Architecture Overview

┌─────────────────────────────────────────────────┐
│           Claude Code CLI                       │
│         (Anthropic's CLI tool)                  │
└─────────────────┬───────────────────────────────┘
                  │
        ┌─────────▼──────────┐
        │  cc-copilot-bridge │  ◄─── This Tool
        └─────────┬──────────┘
                  │
        ┌─────────┴────────────┌─────────────────┐
        │                      |                 │
    ┌───▼────┐         ┌───────▼────────┐   ┌───▼────┐
    │ Direct │         │ Copilot Bridge │   │ Ollama │
    │  API   │         │  (copilot-api) │   │ Local  │
    └────────┘         └────────────────┘   └────────┘
    Anthropic           GitHub Copilot       Self-hosted
    Pay-per-token       Premium requests     Free (offline)
                        quota system

🚀 Quick Start

Installation

Recommended: Package Managers (clean, dependency-managed, easy updates)

brew tap FlorianBruniaux/tap
brew install cc-copilot-bridge
eval "$(claude-switch --shell-config)"

Add to ~/.zshrc: eval "$(claude-switch --shell-config)"

VERSION="1.5.3"  # Check releases for latest
wget https://github.com/FlorianBruniaux/cc-copilot-bridge/releases/download/v${VERSION}/claude-switch_${VERSION}.deb
sudo dpkg -i claude-switch_${VERSION}.deb
eval "$(claude-switch --shell-config)"

Add to ~/.bashrc: eval "$(claude-switch --shell-config)"

VERSION="1.5.3"  # Check releases for latest
wget https://github.com/FlorianBruniaux/cc-copilot-bridge/releases/download/v${VERSION}/claude-switch-${VERSION}-1.noarch.rpm
sudo rpm -i claude-switch-${VERSION}-1.noarch.rpm
eval "$(claude-switch --shell-config)"

Add to ~/.bashrc: eval "$(claude-switch --shell-config)"

Alternative: Script Install (if package managers unavailable)

curl -fsSL https://raw.githubusercontent.com/FlorianBruniaux/cc-copilot-bridge/main/install.sh | bash

Full guides:

• Package Managers - Recommended method
• Quick Start - All installation options
• Install Options - Integration with antigen, oh-my-zsh, etc.

Aliases Included

The installer creates ~/.claude/aliases.sh with these commands:

# Core commands (created automatically)
ccd        # Anthropic API (paid)
ccc        # GitHub Copilot (default: Claude Sonnet 4.6)
cco        # Ollama Local (offline)
ccs        # Check all providers

# Model shortcuts (40+ models)
ccc-opus='COPILOT_MODEL=claude-opus-4-6 claude-switch copilot'
ccc-sonnet='COPILOT_MODEL=claude-sonnet-4-6 claude-switch copilot'
ccc-gpt='COPILOT_MODEL=gpt-4.1 claude-switch copilot'
ccc-grok='COPILOT_MODEL=grok-code-fast-1 claude-switch copilot'
ccc-prod, ccc-dev, ccc-quick, ccc-alt, ccc-private  # semantic shortcuts

See INSTALL-OPTIONS.md for integration with antigen, oh-my-zsh, zinit, etc.

Usage

# Start with Copilot (free via your subscription)
ccc

# Switch models on-the-fly
COPILOT_MODEL=gpt-4.1 ccc
COPILOT_MODEL=claude-opus-4-6 ccc

# Check status
ccs

Visual Examples:

Claude Sonnet 4.6 (Default):

Claude Opus 4.6 (Premium):

GPT-4.1 (OpenAI):

Ollama Offline (Private):

💰 GitHub Copilot Pricing & Limits

Important: Using Claude Code via Copilot consumes your premium request quota. Usage is NOT unlimited.

Current Plans (January 2026)

Model Multipliers

Different models consume different amounts of premium requests per interaction:

Key insight: GPT-4.1 and GPT-4o are free (0x multiplier) on paid plans. Use them for routine tasks to preserve premium requests for Claude/Opus.

Quota Behavior

• Quotas reset on the 1st of each month (00:00 UTC)
• Unused requests do not carry over
• When quota is exhausted, system falls back to free models (GPT-4.1)
• Optional: Enable spending budgets for overflow at $0.04/request

Source: GitHub Copilot Plans

🎨 Features

1. Instant Provider Switching (3 characters)

ccd     # Anthropic Direct API (production)
ccc     # GitHub Copilot Bridge (prototyping)
cco     # Ollama Local (offline/private)

No config changes, no restarts, no environment variable juggling.

Help Menu:

Available commands:

• ccs / claude-switch status - Check all providers health
• claude-switch --help - Full command reference

2. Dynamic Model Selection (40+ models)

# Switch models mid-session
ccc                     # Default: claude-sonnet-4-6
ccc-opus                # Claude Opus 4.6
ccc-gpt                 # GPT-4.1
COPILOT_MODEL=gemini-2.5-pro ccc  # Gemini

# Ollama models
cco                     # Default: devstral-small-2
cco-devstral            # Explicit Devstral
cco-granite             # Granite4 (long context)

3. GPT Codex & Gemini 3 Models (via Unified Fork - RECOMMENDED)

GPT Codex models use OpenAI's /responses endpoint, and Gemini 3 models have thinking support. Both require a fork of copilot-api that combines PR #167 and #170.

⚠️ Important: Codex models are tested and working. Gemini 3 agentic mode is Supported - PR #167 adds thinking support, and tool calling issues have been addressed in fork v1.3.1.

Setup:

# Terminal 1: Launch unified fork (auto-clones if needed)
ccunified

# Terminal 2: Use models
ccc-codex         # gpt-5.2-codex ✅ Tested
ccc-gemini3       # gemini-3-flash-preview ✅ Supported
ccc-gemini3-pro   # gemini-3-pro-preview ✅ Supported

Model Status:

What to test for Gemini 3:

# 1. Baseline (should work)
ccc-gemini3 -p "1+1"

# 2. Agentic mode (uncertain - please report results!)
ccc-gemini3
❯ Create a file test.txt with "hello"

Fork source: caozhiyuan/copilot-api branch 'all' | PR #167 | PR #170

📖 Full guide: docs/ALL-MODEL-COMMANDS.md

5. MCP Profiles System (Auto-Compatibility)

Problem: GPT-4.1 has strict JSON schema validation → breaks some MCP servers

Solution: Auto-generated profiles exclude incompatible servers

~/.claude/mcp-profiles/
├── excludes.yaml       # Define problematic servers
├── generate.sh         # Auto-generate profiles
└── generated/
    ├── gpt.json       # GPT-compatible (9/10 servers)
    └── gemini.json    # Gemini-compatible

6. Model Identity Injection

Problem: GPT-4.1 thinks it's Claude when running through Claude Code CLI

Solution: System prompts injection

~/.claude/mcp-profiles/prompts/
├── gpt-4.1.txt        # "You are GPT-4.1 by OpenAI..."
└── gemini.txt         # "You are Gemini by Google..."

Result: Models correctly identify themselves

7. Health Checks & Fail-Fast

ccc
# → ERROR: copilot-api not running on :4141
#    Start it with: copilot-api start (or scripts/launch-unified-fork.sh)

8. Session Logging

tail ~/.claude/claude-switch.log

[2026-01-22 09:42:33] [INFO] Provider: GitHub Copilot - Model: gpt-4.1
[2026-01-22 09:42:33] [INFO] Using restricted MCP profile for gpt-4.1
[2026-01-22 09:42:33] [INFO] Injecting model identity prompt for gpt-4.1
[2026-01-22 10:15:20] [INFO] Session ended: duration=32m47s exit=0

🏗️ Provider Architecture

🎯 GitHub Copilot Bridge

Use Case: Daily coding, prototyping, exploration

ccc                               # Default: claude-sonnet-4-6
ccc-gpt                          # GPT-4.1 (0x multiplier = free)
ccc-opus                         # Claude Opus 4.6 (3x multiplier)
COPILOT_MODEL=gemini-2.5-pro ccc # Gemini

How It Works:

• Routes through copilot-api proxy
• Uses your Copilot premium request quota (see Pricing & Limits)
• Access to 15+ models (Claude, GPT, Gemini families)
• Best for: Daily development, experimentation, learning

copilot-api Running:

Screenshot: copilot-api proxy server logs showing active connections

Requirements:

1. GitHub Copilot Pro ($10/mo) or Pro+ ($39/mo) subscription
2. copilot-api running locally (copilot-api start or scripts/launch-unified-fork.sh)

🎁 BONUS: Ollama Local (Offline Mode)

Use Case: Offline work, proprietary code, air-gapped environments

cco                                          # Default: devstral-small-2
OLLAMA_MODEL=devstral-64k cco               # With 64K context (recommended)
OLLAMA_MODEL=ibm/granite4:small-h cco       # Granite4 (long context, 70% less VRAM)

How It Works:

• Self-hosted inference (no internet required)
• Free, 100% private
• Apple Silicon optimized (M1/M2/M3/M4 - up to 4x faster)
• Best for: Sensitive code, airplane mode, privacy-first scenarios

Important: Ollama is architecturally independent from Copilot bridging. It's a separate provider for local inference, not related to copilot-api.

⚠️ Critical: Context Configuration

Claude Code sends ~18K tokens of system prompt + tools. Default Ollama context (4K) causes hallucinations and slow responses.

Create a 64K Modelfile (recommended):

mkdir -p ~/.ollama
cat > ~/.ollama/Modelfile.devstral-64k << 'EOF'
FROM devstral-small-2
PARAMETER num_ctx 65536
PARAMETER temperature 0.15
EOF
ollama create devstral-64k -f ~/.ollama/Modelfile.devstral-64k
OLLAMA_MODEL=devstral-64k cco

Recommended Models (March 2026):

SWE-bench measures real-world agentic coding ability (GitHub issue resolution with tool calling, multi-file editing). High HumanEval scores don't guarantee agentic performance.

On the radar (not yet locally runnable):

DeepSeek V4 (released Feb 2026): 1T parameters, 1M context window, Apache 2.0. Top SWE-bench scores but requires 200GB+ RAM even quantized. No runnable distillation confirmed for Ollama yet. Follow DeepSeek releases for Q4 distillations.

Benchmark Sources:

• Devstral-small-2: Mistral AI - 68.0% SWE-bench Verified
• Qwen3-coder: Index.dev - 69.6% SWE-bench Verified
• Qwen3-Coder-Next: dev.to - 42.8% SWE-bench Verified (3B active params)
• GLM-4.7 full: Z.AI - 73.8% (Flash variant "tier lower", no published bench)

Why Devstral despite lower SWE-bench?

• Designed specifically for agentic software engineering tasks (source)
• Native architecture for tool calling vs post-training bolt-on (Qwen3)
• "Best agentic coding" confirmed in practice (CLAUDE.md testing)
• Qwen3 has higher bench but "needs template work" in real usage

⚠️ Models NOT recommended (low SWE-bench despite good HumanEval):

• CodeLlama:13b - 40% SWE-bench (no reliable tool calling)
• Llama3.1:8b - 15% SWE-bench ("catastrophic failure" on agentic tasks)

Requirements:

1. Ollama installed (ollama.ai)
2. Models downloaded (ollama pull devstral-small-2)

Note: Ollama uses GGUF format (universal). For maximum Mac performance with small models (<22B), LM Studio + MLX can be up to 4x faster. However, for models >30B, GGUF becomes more performant. LM Studio is not compatible with claude-switch.

🔄 FALLBACK: Anthropic Direct API

Use Case: Production, maximum quality, critical analysis

ccd

How It Works:

• Official Anthropic API
• Pay per token ($0.015-$75 per 1M tokens)
• Best for: Production code review, security audits, critical decisions

Requirements:

1. ANTHROPIC_API_KEY environment variable
2. Anthropic account with billing

📊 Alternatives

For general multi-provider routing, see @musistudio/claude-code-router (31.9k weekly downloads). For a complete open-source alternative, see OpenCode (48k stars).

cc-copilot-bridge specifically serves Copilot Pro+ subscribers who want to use Claude Code CLI with their existing subscription.

📖 Full Competitive Analysis →

🎬 Real-World Workflows

Workflow 1: Quota-Optimized Development

# Use GPT-4.1 for routine tasks (0x multiplier = doesn't consume quota)
ccc-gpt
❯ Build user authentication flow

# Use Claude Sonnet for complex logic (1x multiplier)
ccc
❯ Design database schema

# Use Anthropic Direct for production review (official API)
ccd
❯ Security audit of auth implementation

Workflow 2: Multi-Model Validation

# Compare approaches across models
ccc-gpt       # GPT-4.1 analysis (free)
ccc           # Claude Sonnet analysis (1x)
ccc-opus      # Claude Opus analysis (3x - use sparingly)

Workflow 3: Offline Development

# Work on proprietary code (airplane mode)
cco
❯ Implement proprietary encryption algorithm
# ✅ No internet required
# ✅ Code never leaves machine

📦 What's Included

🔧 Requirements

• Claude Code CLI (Anthropic)
• copilot-api for Copilot provider
  • Recommended: caozhiyuan/copilot-api v1.3.1 (fork — actively maintained, native Anthropic Messages API, Codex /responses endpoint, gpt-5.4, gemini-3.1)
  • Official: ericc-ch/copilot-api (stalled since Oct 2025, last release v0.7.0)
  • ⚠️ Note: Community patch applied to fix issue #174 (reserved billing header). See TROUBLESHOOTING.md for details.
• Ollama (optional, for local provider)
• jq (JSON processing)
• nc (netcat, for health checks)

📚 Documentation

This Project

• QUICKSTART.md - 2-minute setup
• ALIASES.md - Complete command reference (30+ aliases)
• MODEL-SWITCHING.md - Dynamic model selection guide
• MCP-PROFILES.md - MCP Profiles & System Prompts
• SECURITY.md - Security, privacy, and compliance guide
• OPTIMISATION-M4-PRO.md - Apple Silicon optimization
• TROUBLESHOOTING.md - Problem resolution

Claude Code Resources

• 📖 Claude Code Ultimate Guide - Comprehensive guide to Claude Code CLI
• 🔗 Ultimate Guide Repository - Complete documentation, tips, and best practices

🎯 Who Should Use This?

Primary Audience

• Copilot subscribers who want to use Claude Code CLI with their existing subscription
• Multi-model users who want to compare Claude, GPT, and Gemini responses
• Developers who want a unified interface across multiple AI providers

Secondary Audience

• Privacy-conscious developers who need offline mode for proprietary code (Ollama)
• Teams in air-gapped environments who can't use cloud APIs (Ollama)
• Production users who need Anthropic Direct API for critical analysis

🚀 Version

Current: v1.7.0

Changelog: See CHANGELOG.md

⚠️ Risk Disclosure

Terms of Service Considerations

This project uses copilot-api, a community tool that reverse-engineers GitHub Copilot's API.

Important disclaimers:

1. Not officially supported: copilot-api is not endorsed by GitHub, Microsoft, Anthropic, or any AI provider
2. ToS risk: Using third-party proxies to access Copilot may violate GitHub Copilot Terms of Service
3. Account suspension: GitHub reserves the right to suspend accounts for ToS violations "at its sole discretion" without prior notice
4. API changes: This tool may stop working at any time if providers change their APIs
5. No guarantees: The authors provide no warranty and accept no liability for account suspension or service interruption

Documented Risks

Community reports indicate that:

• Accounts using high volumes through third-party proxies have been suspended
• Suspensions may affect your entire GitHub account, not just Copilot access
• GitHub does not provide a public definition of "excessive usage" or "abuse"

Recommendations

Source: GitHub Terms of Service - API Terms

Related Projects

Enhance your Claude Code workflow:

• Claude Code Ultimate Guide - Documentation and best practices
• ccboard - Dashboard for monitoring and analytics
• RTK - Token reduction proxy
• Claude Cowork Guide - Desktop automation guide

More: florian.bruniaux.com

📖 Credits

• copilot-api: ericc-ch/copilot-api - The bridge that makes this possible
• Claude Code: Anthropic - The CLI tool we're enhancing
• Ollama: ollama.ai - Local AI inference

📄 License

MIT

🔗 Related Projects

By the Same Author

• 📖 Claude Code Ultimate Guide - Comprehensive guide to mastering Claude Code CLI
  • Complete documentation and best practices
  • Tips & tricks for productivity
  • MCP server integration guides
  • GitHub: claude-code-ultimate-guide

Community Tools

• copilot-api - GitHub Copilot API proxy (core dependency)
• Ollama - Local AI inference platform
• awesome-claude-code - Curated list of Claude Code resources