DeepSeek Harness
A layer on top of other agents, not a replacement for them.
Verified: Sep 12last change: Sep 20next price check: Dec 11
Facts with sources
- Released August 13; within three weeks, about 165 plugins had piled up in the default install, and some of the published ones are malicious. Source (in Russian)
- Deploying it on a clean server took the guide’s author about 25 minutes, with a second AI on hand to suggest commands. Source (in Russian)
- Can call Claude Code and Codex as subagents. Source (in Russian)
Strengths
- MIT licence; one npx command opens a browser dashboard on port 3080; any provider, including Bedrock, Vertex, Azure or a custom URL.
- The trajectory log shows every call, the system prompt, reasoning and token spend per step, and a session can be exported for review.
- Nate Herk, same Opus model: a YouTube analytics Excel in about 3 minutes vs 17 in Claude Code, a wiki search in 50 seconds vs 5+ minutes.
- Reads AGENTS.md, CLAUDE.md and Claude skills with no changes; can call Claude Code and Codex as subagents.
Weaknesses
- Officially a preview promising breaking changes: for Nate Herk creator mode kept reverting, compaction lost context, his mouse locked once.
- DeepSeek's own models cannot see images: BridgeMind's V4 Pro failed a window-overlap bug given a screenshot, Grok 4.6 fixed it in one shot.
- Unvetted plugins: a community audit found only 17 of 122 catalogued plugins install and 10 pass checks; AI RANEZ warns some carry malware.
- Discussion 461 (Aug 13): an agent in full-access mode wiped a user's entire home directory (Edward Grishin).
Example from practice
An Excel report on a YouTube channel's analytics
Nate Herk sent the same vague prompt to DeepSeek Harness and Claude Code, both on Opus 5: find the recently pulled channel analytics and build an Excel sheet from them. The Harness came back in about 3 minutes, Claude Code in 17. The Harness version was shorter and easier to read: colour-coded views and engagement, top and bottom 10, monthly trend, tag analysis, though its charts had no labels. Claude Code's was wordier but deeper: all 467 videos, a top 50, more tabs and metrics. He would hand a client the Claude Code file; on a research report using his own skill, the Harness produced 4,400 words with 14 load-bearing sources against Claude Code's 5,000 words and 26.
What reviewers say
Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.
- 100 Hours Testing Deepseek Harness vs. Claude Code. What You Need to Know. · Nate Herk | AI Automation, Aug 23The reviewer's view: this is not a free Claude Code but a different harness, faster and endlessly customisable, yet it does not replace Claude Code and he trusts Claude Code's output more.
- Vibe Coding With DeepSeek V4 Pro & DeepSeek Harness. Does It Beat Claude Code? · BridgeMind, Aug 16The reviewer spent $5.87 on 114M tokens for the whole video and rates the price 10 of 10, but will not adopt it: the model cannot see the screen and tasks ran 90 minutes.
- DeepSeek Harness – Конец эпохи Claude Code и Codex? · ZProger [ IT ], Aug 18The reviewer is enthusiastic about 'everything is a plugin' and the reasoning log; he shows the harness writing a plugin for itself and installing a community terminal UI.
- ПОЛНЫЙ ГАЙД по DEEPSEEK HARNESS: установка и использование | Бесплатная замена Claude Code · Академия нейросетей Сергея Нагорного GORA ACADEMY, Aug 26The reviewer would only install it on a disposable 4 GB server, with a second AI dictating commands, and it took about 25 minutes; the program is free, the compute is not.
- Почему все говорят про DeepSeek harness? | Стоит переходить?! · AI RANEZ, Aug 27The reviewer will not switch yet, it is too raw; the two real differences from Claude Code and Codex are the trajectory view and customisation, and web search needs its own key.
- DeepSeek Harness, Grok Build, OpenWorker, qm. ОБЗОР 4 HARNESS. ЧТО Я НАШЕЛ В КОДЕ, ВАС УДИВИТ! · Эдвард Гришин | AI и автоматизация бизнеса, Sep 1The reviewer advises waiting for a stable release: the code is open but development is closed, issues and pull requests are disabled, and there is no Russian UI and no terminal.
Who it is for
- Engineer who already runs several agents and wants a layer over them
Models in this agent
Score is from our manager-task benchmark; models without a score are not measured yet.
What changed
- DeepSeek API: Chinese national holidays now count as discounted off-peak days, on top of weekends →
- DeepSeek V4.1 Flash released: new architecture, 763B parameters, open weights, 1M context, native multimodality; a radically smaller KV cache makes long agent sessions cheaper and faster, and the API is another 50% cheaper off-peak →
- DeepSeek V4 Flash Vision, an experimental multimodal version: reads screenshots and diagrams, files up to 64 MB →
- DeepSeek Harness released: everything is a plugin, it calls Claude Code and Codex as sub-agents
- DeepSeek V4 Flash 0731: the update beat V4 Pro on benchmarks with 10 times fewer active parameters, open weights, $0.14 per million tokens →
- DeepSeek released V4 Pro and V4 Flash: analysing 10 thousand lines of code in Flash costs under a cent →
How to brief an agent like this
The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.