WayToClawEarn
High impactHacker News

DeepClaude open source project explodes: Use DeepSeek V4 to run Claude Code, the cost is only 1/17

The open source project DeepClaude allows developers to use DeepSeek V4 Pro instead of Anthropic to run the Claude Code agent loop, reducing the cost to 1/17. Earned 179 points on Hacker News in just 4 hours and supports real-time switching of multiple backends.

WayToClawEarn EditorialPublished May 4, 2026Updated Aug 8, 2026

Editorial review of public sources · AI-assisted drafting. How we work · Original source

Core conclusion

On May 4, 2026, the open source project DeepClaude exploded on Hacker News, receiving 179 votes and 86 comments within 4 hours. This project uses environment variable hijacking technology to allow Claude Code's autonomous Agent loop to directly call DeepSeek V4 Pro for inference. The API cost is reduced from Anthropic's $15/M output tokens to DeepSeek's $0.87/M, saving 94% (about 17 times).

Key Points

  • Event Time: 2026-05-04, HN Home Page #5
  • Core changes: Claude Code's Agent loop (file editing, bash execution, sub-Agent generation) remains unchanged, only the back-end inference model is replaced
  • Price Difference: DeepSeek V4 Pro $0.87/M output vs Anthropic $15/M output
  • Practicality: Not a toy project, 80% of daily tasks are as effective as Claude Opus

What is DeepClaude

DeepClaude is a minimalist open source tool (the core consists of only two script files) that routes API calls to DeepSeek's Anthropic-compatible endpoints by setting Claude Code's environment variables ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, and ANTHROPIC_MODEL.

It only does three things: change the route, record the cost, and automatically restore the original settings next time.

Technical principles

HierarchyComponentFunction
Terminal layerClaude Code CLITool loop, file reading and writing, bash execution, git operation - unchanged
Routing layerdeepclaude.sh / .ps1Set environment variables, start agents, cost tracking — New
Inference LayerDeepSeek V4 ProReplacement Anthropic model handles all inference requests — Replacement

Price comparison: really 17 times?

Usage scenariosAnthropic Max Plandeepclaude (DeepSeek)Savings
Light usage (10 days/month)$200/ months~$20/ months90%
Moderate usage (25 days/month)$200/ months~$50/ months75%
Automatic Agent Cycle$200/ Month~$80/ Month60%

DeepSeek V4 Pro's automatic context caching significantly reduces the cost of Agent cycles - system prompts and file contexts after the first round are priced in $0.004/M tokens (original price $0.44/M), a reduction of 120 times.

Core functions

1. Multiple backend support

DeepClaude not only supports DeepSeek, but also supports OpenRouter ($0.44/M input, lowest latency), Fireworks AI (fastest inference speed), and original Anthropic (switch back at any time when dealing with complex problems).

2. Live Switching

Switch between inference backends on the fly with slash commands such as /deepseek, /anthropic, etc. without restarting your Claude Code session. Under the hood is a local proxy (localhost:3200) that intercepts all API calls and provides a control endpoint:

  • /_proxy/mode → Switch backend
  • /_proxy/status → current backend + running time
  • /_proxy/cost → Token usage + cost saving statistics

3. Remote control (Remote)

Supports opening Claude Code sessions in any browser through deepclaude --remote, inference still uses DeepSeek, and the WebSocket bridge of remote sessions uses Anthropic infrastructure.

Limitations and Notes

FunctionStatusDescription
Image/Visual Input❌ Not SupportedDeepSeek compatible endpoints do not support images
Parallel tool invocation❌ DowngradeTools are executed serially one by one
MCP server tool❌ Not supportedCompatibility layer is not supported yet
Complex reasoning (20% of scenes)⚠️ DowngradeIn a few scenes it is still recommended to switch back to Anthropic

AI cost comparison chart

Implications for AI content producers

The significance of DeepClaude is not only to save money, it also sends an important signal:

**When you are no longer constrained by API cost, the usage pattern of AI Agent workflows changes completely. **

In the past, AI content producers often had to weigh "whether each call is worth it" when orchestrating automated workflows. When the cost drops to 1/17, a large number of scenarios that were rejected due to cost in the past become feasible:

  • Batch rewriting and polishing
  • Multiple rounds of review cycles (Agent repeatedly checks→modifies→checks again)
  • Complex multi-step orchestration process

Combining Claude Code's autonomous Agent capabilities with DeepSeek V4 Pro's low-cost inference, content producers can maintain high-frequency AI Agent collaboration at a cost of less than $100/ months.

Related extended information

Tool entry

Tools mentioned in the text: DeepSeek, Claude, Claude Code, OpenRouter, Anthropic

Internal link guidance

View source →

Disclaimer: this site shares educational insights only, for inspiration and reference. No outcome guarantee; external execution and decisions are your own responsibility.