DeepClaude — Claude Code Multi-Backend Proxy
Run Claude Code's autonomous agent loop with DeepSeek V4 Pro, OpenRouter, or any Anthropic-compatible backend. Same UX, 17x cheaper.
Evidence shows: scripts set environment variables and start a proxy, but least privilege is not explicitly addressed; users must manually set API keys, but proxy auto-start may not have explicit confirmation; data flow diagram clearly shows proxy path, but logging or data retention is not mentioned; API keys are handled as environment variables, but encryption or secure storage is not mentioned; dependencies are not listed, so security cannot be assessed; proxy may modify system environment variables, but impact scope is not stated; original settings are restored on exit, but no rollback mechanism is provided; author identity is unverified, but README and LICENSE provide attribution. Deductions: lack of least privilege, user confirmation, sensitive data handling details, dependency security, external impact scope, rollback mechanism, and source verification.
Evidence shows: README is consistent, but no tests or verification are provided; dependencies are not listed, so availability cannot be confirmed; error handling is not detailed. Deductions: lack of test evidence, unverified dependency availability, unclear error messages.
Evidence shows: clear target audience of developers, scenarios include terminal and VS Code; capability boundaries clearly list supported and unsupported features; trigger methods are clear (command line and slash commands); environment supports Windows, macOS, Linux, but specific version requirements are not mentioned. Deductions: environment fit lacks version compatibility details.
Evidence shows: README structure is clear, including installation, usage, configuration; installation steps are detailed; naming is stable but versioning is not mentioned; examples and FAQ are provided; known limitations are clearly listed; MIT license is complete; no version history or changelog; maintenance responsibility is not clear. Deductions: lack of versioning, changelog, and maintenance responsibility.
Evidence shows: output is terminal commands and proxy responses, format is clear; cost savings and performance comparisons are provided, marginal value is evident; cost-benefit analysis is detailed, but no actual test data is provided. Deductions: lack of actual run verification.
Evidence shows: claims in README lack external sources; no cross-validation; facts and inferences are mixed, e.g., performance comparisons lack sources. Deductions: lack of traceability, cross-validation, and fact-inference separation.
- The proxy intercepts all API calls, potentially introducing man-in-the-middle risks; ensure the proxy source is trustworthy.
- API keys are stored as environment variables without mention of encryption or secure storage, posing a leak risk.
- Dependencies are not listed, so supply chain security cannot be assessed.
- Author identity is unverified; use with caution.
What does this agent do, and when should you use it?
DeepClaude is a tool that lets you use Claude Code's autonomous agent loop with cheaper model backends. It redirects Claude Code's API calls to DeepSeek V4 Pro, OpenRouter, or Fireworks AI by setting environment variables like ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, while preserving full tool-use capabilities (file editing, bash execution, subagents). The project includes shell/PowerShell launchers (deepclaude.sh/deepclaude.ps1), a local proxy (localhost:3200) for live backend switching and cost tracking, and remote control (--remote) for browser access. It supports switching backends mid-session via slash commands, CLI flags, or VS Code tasks.
DeepClaude launches Claude Code with environment variables pointing to an alternative backend. It provides CLI commands like deepclaude --backend or (OpenRouter) and deepclaude --cost (pricing comparison). The local proxy at localhost:3200 intercepts API calls and enables live switching via /_proxy/mode endpoint, with cost tracking via /_proxy/cost returning token usage and savings. Remote control (--remote) starts a proxy that routes model calls to DeepSeek while maintaining Anthropic's bridge WebSocket. The scripts validate API keys (e.g., DEEPSEEK_API_KEY) and support subagent model mapping.
- Individual developers who want Claude Code's autonomous agent capabilities at a lower cost using DeepSeek V4 Pro.
- Teams looking to save on heavy agent usage by moving from a flat $200/month Claude plan to pay-per-token backends.
- Users who need to switch models mid-session, e.g., from DeepSeek to Anthropic for complex reasoning tasks, without restarting the CLI.
- Developers who want to control a Claude Code session remotely from a browser or mobile device, routing model calls to a cheaper backend.
- Users who want to monitor token usage and cost savings via the proxy's /_proxy/cost endpoint.
- Teams that prefer using OpenRouter's multi-model routing rather than a single model provider for their agent workflows.
What are this agent's strengths and limitations?
- Significant cost reduction: DeepSeek output at ~$0.87/M tokens vs Anthropic's $15/M, roughly 17x cheaper.
- Preserves Claude Code's full functionality, including file editing, bash execution, and subagent spawning.
- Supports live switching between multiple backends (DeepSeek, OpenRouter, Fireworks) via CLI, slash commands, or VS Code tasks.
- Built-in cost tracking via /_proxy/cost endpoint to quantify savings.
- Provides a remote control option to use in a browser while routing model calls to non-Anthropic backends.
- Text-only: image/vision input is not supported (DeepSeek's Anthropic endpoint lacks image support).
- MCP server tools are not supported through the compatibility layer.
- Anthropic's cache_control is ignored; only DeepSeek's automatic caching is used.
- Complex reasoning (about 20% of work) still requires Anthropic Opus, potentially offsetting savings.
- Installation requires manual PATH modifications and symlinks, which may be a barrier for non-technical users.
How do you install or deploy this agent?
On Windows, copy deepclaude.ps1 to a PATH directory or add the repo directory to PATH. On macOS/Linux, chmod +x deepclaude.sh and symlink to /usr/local/bin/deepclaude. Before installing, obtain API keys: sign up at platform.deepseek.com, add $5 credit, and set DEEPSEEK_API_KEY (setx on Windows, export on macOS/Linux). Optionally set OPENROUTER_API_KEY or FIREWORKS_API_KEY. Node.js 18+ is required for the proxy.
How do you use this agent?
Run deepclaude in your terminal to launch Claude Code with DeepSeek V4 Pro. Use --backend flag to select backend: deepclaude --backend or (OpenRouter), --backend fw (Fireworks AI), --backend anthropic (Anthropic). To switch live during a session, add custom slash commands (e.g., ~/.claude/commands/deepseek.md) or run deepclaude --switch ds. For remote control, use deepclaude --remote. Check cost with curl http://127.0.0.1:3200/_proxy/cost.
How does this agent compare with similar options?
The repository explicitly compares Anthropic's Claude Opus against alternative backends (DeepSeek, OpenRouter, Fireworks) in terms of pricing and intelligence. It positions itself as a cheaper alternative for routine tasks, reserving Claude Opus for complex reasoning.