DeepSeek Harness - DeepSeek

DeepSeek Harness

DeepSeek · Terminal & CLI Agents · Updated Aug 13, 2026

DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI, providing the engineering layer (context management, tool calling, file editing, terminal execution) that turns DeepSeek models into coding agents.

Recent Updates

DeepSeek Harness was first mentioned in the July 31, 2026 changelog as 'to be released soon'. A community version (HenryZ838978/deepseek-harness) was published on May 9, 2026, with Python library, CLI, and MCP server. DeepSeek formed a dedicated Harness team in May 2026 to build a Code Harness product.

As of Aug 13, 2026

Key Features

Context management and memory
Tool calling and file operations
Terminal execution
Self-correction with feedback loops
MCP server support
Anthropic SKILL.md integration
OpenAI-compatible API

Who It's For

✓ Best For

Developers and enterprises seeking a low-cost, open-source alternative to Claude Code for autonomous coding tasks, especially those already using DeepSeek models.

✗ Not Ideal For

Teams requiring mature IDE integration, extensive Skills ecosystem, or polished MCP tooling, as DeepSeek Harness is still in early stages and lacks the ecosystem depth of Claude Code.

Pricing

Open Source

DeepSeek Harness is open-source and free to use. The underlying DeepSeek V4 models are available via API with usage-based pricing, e.g., $0.14 per million input tokens (cache miss) for V4 Flash and $0.435 for V4 Pro, with off-peak discounts.

Verdict

DeepSeek Harness is a promising open-source agent framework that leverages DeepSeek's cost-effective models, but it is not yet officially released and lacks the maturity and ecosystem of established tools like Claude Code.