Published: 2026-08-23

DeepSeek Harness vs Claude Code After a Week of Real Use

DeepSeek shipped an open-source, free agent harness whose pitch is that everything is a plugin — tools, skills, sessions, sandboxes, storage, the agent loop, even the UI. Nate Herk ran it for a week against Claude Code on identical prompts with the same model underneath, and the timing gaps are large enough to be worth knowing about: 50 seconds versus five-plus minutes on a wiki search, three minutes versus seventeen on the same deliverable. He is also clear about what it is not — a free Claude Code — and honest about the preview bugs.

Source video

"100 Hours Testing Deepseek Harness vs. Claude Code. What You Need to Know." by Nate Herk — Watch on YouTube →

Getting It Running

Install: he handed the GitHub link to Codex and asked it to set the harness up. It opens on a localhost URL and runs entirely locally.
Add a model key: Settings → Models → paste a key. He used an OpenRouter key, which makes every OpenRouter model selectable from the picker.
Billing: the harness is free, inference is not. You cannot use a Claude or Codex subscription by default — it is per-token API billing. He has seen plugins that claim to bridge a subscription but has not tested them.
Existing config just worked: he dropped it into his existing setup without changing anything, and it picked up his Claude skills and context files as-is.

The Four Modes

Mode	What it is	When he reaches for it
Standard	Full coding agent — file editing, shell, file access, web search. On the first prompt it does a context injection: reads `AGENTS.md`, `CLAUDE.md` and `CLAUDE.local.md`, then the DeepSeek system prompt, then the skill catalog.	Default. Anything where the agent needs to know who you are and what your project is.
PTC	Parallel/multi-chain execution for large tasks.	Big jobs with many independent parts. He reports noticeably better output than standard mode on this shape of work.
Minimal	Two-tool coding agent with a persistent bash session. Skips the context injection entirely — no `CLAUDE.md`, no skill catalog.	Quick one-off tasks. Fast, cheap, efficient — but the agent will not know your project.
Creator	For building custom agent presets, writing plugins, and changing the UI.	When you hit a limitation and want to fix the harness itself rather than work around it.

He demonstrates the standard/minimal difference directly: the same "hi, who am I" prompt triggers a full context read in standard mode and nothing at all in minimal.

The Side-by-Side Results

All three comparisons used identical prompts fired at the same time. The Excel and research tests used the same model (Opus 5) in both harnesses, so the difference is the harness, not the model.

Task	DeepSeek Harness	Claude Code	His verdict
Find a half-remembered project in a large personal wiki	~50 seconds	5+ minutes	DeepSeek, consistently and by a wide margin across repeated runs
Pull YouTube analytics and build an Excel breakdown	~3 minutes — colour-coded columns, top/bottom 10 charts, easy to read	~17 minutes — more tabs, all 467 videos, deeper drill-down, wordier	Claude Code for quality, DeepSeek for legibility and speed
Run his STORM research skill on one topic	~4,400 words, 14 load-bearing sources (22 found), practical and relatable, occasionally overconfident	~5,000 words, 26 sources, more scientific, more conservative with claims	Claude Code — the one he'd hand to a client

Both research reports followed the skill's structure correctly, which is the interesting part: the same skill, run by the same model, produced measurably different depth depending only on the harness wrapped around it.

The Trajectory View

The feature he singles out as genuinely novel. Clicking into the trajectory panel shows granular intermediate steps for the whole session in one place — user prompt, context that came in, every action taken — and the session log is downloadable. His stated use for it is closing the loop: look at what went wrong in a session, then say "build a plugin so that doesn't happen next time," which is only practical because the harness is modifiable.

Gotchas & Caveats

Vet third-party plugins before installing. His own warning: an open plugin ecosystem means arbitrary code from strangers. Have Claude Code or Codex review any plugin before you install it. This is the same threat model as skills — see our Security Center.
It is labelled a developer preview and behaves like one. He hit compaction bugs and context regressions in long sessions, a creator-mode button that flipped straight back on click, and one occasion where the app locked his mouse to a small square of the screen.
Cheap tokens are not automatically cheap runs. His framing: judge cost-to-completion, not cost-per-token — an inefficient loop on a cheap model can lose to an efficient loop on an expensive one.
The DeepSeek models can't see. He built a site with DeepSeek V4 Flash that looked fine but contained a visual error Claude Code or Codex would have caught, because those can screenshot, look, and iterate. Match the model to the task — design work needs vision.
Skills are harness-specific in practice. The skills he tested were written for Claude Code, and he expects a skill authored for DeepSeek Harness to perform substantially better in it.
It does not replace Claude Code. His conclusion, stated plainly: this is not a free Claude Code, it is a different harness. If you want Opus, the Claude subscription gives far better value than paying per token for it here.

Key Takeaways

The harness is free and open source; the inference is not — expect per-token API billing unless a plugin bridges your subscription.
Speed is the headline result, and it held up across repeated identical-prompt tests with the same model on both sides.
Claude Code produced the more thorough, more conservative, more client-ready output in both quality comparisons.
Minimal mode's lack of context injection is a feature, not an omission — but it means the agent won't know your project, so don't use it for project work.
Full customisation is the real differentiator: with a closed harness you can swap the model but not the agent loop, the tools, or the UI.
Existing CLAUDE.md, AGENTS.md and skill setups transfer without modification, so trying it costs almost no setup time.