DeepSeek Harness v0.1: The Open-Source Agent Framework That Broke GitHub Records

On August 13, 2026, DeepSeek made a pair of coordinated moves — open-sourcing a new agent framework while shipping a major model update and announcing API price increases — sending a mixed signal about openness versus monetization.

DeepSeek Harness v0.1 — Open-Source Agent Framework

DeepSeek released Harness v0.1 as a developer preview under the MIT license on GitHub at github.com/deepseek-ai/deepseek-harness . The framework is built on the Cordis meta-framework with an "everything is a plugin" architecture — models, tools, skills, sessions, sandboxes, filesystems, agent loops, orchestration, scheduling, and UI are all swappable plugins .

DeepSeek positioned Harness as an open-source rival to Anthropic's Claude Code and OpenAI's Codex/Codex CLI . The repo's GitHub traction was explosive: it passed ~24,000 stars within 1.5 hours and 28,000 within 3 hours of launch . As of the latest snapshot, the repo showed 32.2k stars and 2.4k forks with 19 contributors .

You can launch Harness via

npx @deepseek-ai/dsh web

DeepSeek-V4-Pro-0813 — Model Upgrade

DeepSeek-V4-Pro-0813 went general availability on August 12–13, 2026, replacing the preview build behind the deepseek-v4-pro API endpoint . It is a 1.6 trillion total parameter Mixture-of-Experts model (49 billion active per token) with a 1-million-token context window and up to 384K output tokens . The 0813 build ships with enhanced agent capabilities and native OpenAI Responses API compatibility . V4 Pro's open weights are also MIT-licensed .

Organizational Push — The Harness Team

DeepSeek began assembling a dedicated "Harness team" in May 2026, internally targeting Claude Code and Codex . Two Beijing-based roles were posted: Harness Product Manager and Harness R&D Engineer . The team's charter was to build "Code Harness" (sometimes referred to internally as "DeepSeek Code") as a desktop Agent product bridging DeepSeek's model capabilities into a full agent experience .

API Price Hikes — Tension with Openness

Effective August 16, 2026, DeepSeek introduced peak/off-peak pricing tiers for the V4 Pro API . Output pricing rises from the preview rate of $0.87 per million tokens to $1.98 off-peak and $3.96 during peak hours — a 2.3× to 4.6× increase . This reverses earlier price cuts and creates a strategic tension: the agent stack (Harness) becomes more open (MIT license, fully modular, self-hostable), while the hosted API becomes considerably more expensive, especially during peak hours .

Strategic Contrast

Competitors (Anthropic with Claude Code, OpenAI with Codex CLI) offer hosted, vertically integrated agent experiences where the model, tooling, sandbox, and runtime are tightly coupled and proprietary . DeepSeek's bet is the opposite: open-source the entire agent harness under MIT, let developers swap in any model adapter, any sandbox, any UI — while monetizing the hosted inference tier at premium rates . Developers who find the API too expensive can self-host the open-weight V4 Pro model and run it through Harness locally .

In short: August 13, 2026 was DeepSeek's most consequential single-day release — open-sourcing its agent infrastructure to compete with Claude Code/Codex, GA-launching its 1.6T-parameter flagship, and simultaneously jacking up hosted API prices by 2–4×, creating a clear "open the stack, raise the toll" strategy.