deepclaude

// Overview

deepclaude keeps the Claude Code CLI exactly as it is — the tool loop, file editing, bash execution, subagent spawning — and changes only where the model calls go. It works by setting the environment variables Claude Code reads to decide its endpoint (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN and the per-tier model names), launching the CLI with them, and restoring your original settings when the session ends. Nothing is written permanently, so a normal `claude` run is unaffected.

Four backends ship out of the box: DeepSeek (the default), OpenRouter, Fireworks AI and Anthropic itself, each selected with a `--backend` flag. The project's own comparison table lists DeepSeek and OpenRouter at $0.44 per million input and $0.87 per million output tokens against Anthropic's $3.00 / $15.00, which is where the README's "17x cheaper" headline comes from.

A second mode runs a local proxy on port 3200 that intercepts `/v1/messages` and passes everything else through to Anthropic. Because the proxy exposes a control endpoint, you can flip the active backend from inside a running session — the README wires it to `/deepseek`, `/anthropic` and `/openrouter` slash commands — and read token usage and accumulated savings from `/_proxy/cost`. The README is explicit about the limits of the compatibility layer: image input, MCP server tools and Anthropic-style `cache_control` prompt caching do not carry over.

// What it does

• Per-session environment shim: launches Claude Code against a different backend without changing your permanent configuration
• Four backends behind one flag — DeepSeek, OpenRouter, Fireworks AI and Anthropic — plus `--status`, `--cost` and `--benchmark` helpers
• Local proxy on port 3200 with a control endpoint, so you can switch models mid-session from a slash command instead of restarting
• Cost tracking: `/_proxy/cost` returns tokens, spend and the Anthropic-equivalent price for the same traffic
• `--remote` splits traffic so Claude Code's browser remote-control bridge stays on Anthropic while model calls go to the cheap backend
• Shell scripts for macOS/Linux and PowerShell for Windows, plus VS Code and Cursor terminal-profile snippets

// Getting started

deepclaude wraps an existing Claude Code install, so set up the CLI first. The steps below follow the README's two-minute quick start on macOS/Linux; Windows uses `deepclaude.ps1` and `setx` instead.

echo 'export DEEPSEEK_API_KEY="sk-your-key-here"' >> ~/.bashrc
source ~/.bashrc

chmod +x deepclaude.sh
sudo ln -s "$(pwd)/deepclaude.sh" /usr/local/bin/deepclaude

deepclaude                      # DeepSeek
deepclaude --backend or         # OpenRouter
deepclaude --backend anthropic  # back to Claude

curl -sX POST http://127.0.0.1:3200/_proxy/mode -d "backend=deepseek"
curl -s http://127.0.0.1:3200/_proxy/cost

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run long autonomous coding loops on a cheap model and reserve the expensive one for the hard problems
• Try a Claude Code workflow against several providers and compare latency with the built-in `--benchmark` mode
• Keep an audit of what an agent session actually cost, and what the same traffic would have cost on Anthropic
• Work from a phone or tablet with `--remote` while the model calls run on a lower-cost backend

// How deepclaude compares

deepclaude alongside other open-source gateways & routing tools AI/TLDR tracks, ranked by GitHub stars.