~ / agents / deepseek-harness

DeepSeek Harness

DeepSeek's MIT-licensed agent harness, a developer preview built on a 'everything is a plugin' Cordis kernel, launched the same day as DeepSeek V4-Pro on the API.

VendorDeepSeek
FormOpen-source agent harness (CLI dsh + local web UI)
LaunchedAugust 13, 2026 (developer preview)
ARscorenot scored
Pricing tierFree (MIT); pay-per-token API
Coding modelsProvider-agnostic: DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, custom
Open sourceYes (MIT)

Official links

Outbound, for trial and verification.

Official sitehttps://github.com/deepseek-ai/deepseek-harness
Pricinghttps://platform.deepseek.com/
English interface / changeloghttps://github.com/deepseek-ai/deepseek-harness

Capabilities

What it ships, in its own terms.

The model adapter, tool registry, session log and even the agent loop are swappable plugins; a local web UI runs at 127.0.0.1:3080 via npx @deepseek-ai/dsh web, with profile and headless CLI modes. It is not locked to DeepSeek models.

Use cases

Scenario coverage (presence, not benchmark performance).

ScenarioLocal filesDocs / PPTExcel reportsAnalysisWeb searchMini-programsWeb building
DeepSeek Harness

✓ supported · ◐ partial or via a sibling product · ? unclear · — not in scope. Presence, not benchmark performance.

Users

Subscriber and active-user counts, source-labelled.

~169K GitHub stars and 18.1K forks one week after releaseSource: Marketplace (GitHub)

Third-party review

Independent coverage, not vendor marketing.

Third-party editorial: no editorial score; V4-Pro self-reported DeepSWE +49.9 / Cybergym +30.6 (vendor, not reproduced)

Coding ability (benchmark): not scored — no DeepSWE or Terminal-Bench measurement.

VentureBeat called it 'an open-source rival to Claude Code'; the README warns of compatibility-breaking changes. V4-Pro's self-reported DeepSWE (+49.9) and Cybergym (+30.6) gains are vendor figures, not independently reproduced.

Source: Press (VentureBeat) / Vendor-reported

Compare further

The reference agents and the scored Chinese entry.

China big-tech hub

all entries vs Codex and Claude Code

comparison

Claude Code

the architecture leader

score 86

ChatGPT Codex

the reference OpenAI terminal agent

score 84

Qwen CodeOS

the scored open-source Alibaba CLI

score 45