DeepSeek Harness vs Claude Code: The Open-Source Agent Runtime vs the Polished Incumbent

Deep​Seek Harness is not a model — it is an agent runtime, and that distinction is the whole comparison. Claude Code is Anthro​pic's polished, closed-source coding agent that just works in your terminal. Deep​Seek Harness (DSH) is an open-source (MIT) agent framework that Deep​Seek released on August 13, 2026, and it connects any model to tools, a local terminal sandbox, and a browser UI. If you want a coding agent you can run today without thinking, Claude Code wins. If you want an open, model-agnostic runtime you can rebuild, and you can tolerate a rough v0.1 preview, DSH is the more interesting bet — and it does not even require Deep​Seek models.

The launch numbers say people think this is a real rivalry: DSH hit roughly 50,000 GitHub stars within its first 12 hours and around 126,000 by August 16 (GitHub repo page, fetched August 16, 2026). That is the fastest agent-runtime adoption on record, and it is squarely aimed at Claude Code. But because DSH is a harness rather than a model, the honest question is not "which is smarter" — it is which layer you want to own: the ready-made product, or the machinery you can take apart.

What Deep​Seek Harness actually is

Deep​Seek's internal formula, stated publicly, is Model + Harness = Agent. The model is the brain; the harness is everything that makes it work as an agent: context and memory management, tool calling, task planning, file editing, terminal execution, error feedback, and judging when a task is finished. Claude Code bundles one vendor's model with a purpose-built harness. DSH unbundles them — the harness is a standalone product, and the model is a swappable plugin.

That architecture is DSH's defining idea: "Everything is a plugin." Built on the Cordis microkernel — by Cui Tianyi, the author of the Koishi chatbot ecosystem and a former nine-year Jane Street engineer — DSH treats model adapters, tools, skills, sessions, the sandbox, storage, approval policies, the agent loop itself, and even its own UI as replaceable plugins. A default install loads 100+ plugins. There is no privileged core to fork; you reconfigure instead.

DSH ships four presets. Standard is the full coding agent. PTC (Programmatic Tool Calling) has the model write TypeScript that batches many tool calls into one deterministic program, so intermediate data never round-trips through the model's context — a serious token saver. Minimal exposes only a shell and a file editor, for benchmarking. Creation lets you build your own presets and plugins, in memory, at runtime. You launch the whole thing with code>npx @deepseek-ai/dsh web/code>, which opens a local web UI at http://127.0.0.1:3080 — the browser is a shell; every session and log stays on your machine.

Three features explain the enthusiasm. First, Trajectory: an append-only session log of everything the model saw — system prompts, reasoning, tool calls, sub-agent scheduling — that you can replay, fork, and audit like a black box. Second, MCP support on both the client and server side, so it talks to the same tool ecosystem as Claude Code. Third, 40+ model providers, including Deep​Seek, Anthro​pic, and Open​AI — you can put Claude models inside DSH, or Deep​Seek models inside Claude Code-compatible tooling, or anything else.

Feature by feature

The table below is the honest snapshot. The two tools agree on the big surface (MCP, sub-agents, session history, model-agnostic plumbing at the protocol level) and differ on the layer they ask you to own.

The differences that matter most:

The price question changed on August 16

Cost is where DSH was supposed to win outright — and it still does per token, but the gap just narrowed. Deep​Seek moved its API to peak/off-peak billing effective 16:00 UTC on August 16, 2026 (peak hours 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak). Under the new schedule, DeepSeek V4 Flash output costs $0.66 per million tokens off-peak and $1.32 at peak, up from $0.28; DeepSeek V4 Pro output goes from $0.87 to $1.98 off-peak and $3.96 at peak. Deep​Seek's own API docs list both the legacy and new rates side by side.

Against Anthro​pic's published API rates — Claude Haiku 4.5 at $1 in / $5 out, Claude Sonnet 4.6 at $3 / $15, Claude Opus 4.8 at $5 / $25 per million tokens — DeepSeek V4 Pro is still roughly 4–6 times cheaper per token even at the new peak output price, and DeepSeek V4 Flash is roughly 10–20 times cheaper than Sonnet 4.6. Cache-hit input is nearly free on both sides, which matters because agent loops are cache-heavy. The direction of the price war, though, flipped: Deep​Seek raised prices while Anthro​pic has been cutting or holding its agent rates.

The counterweight is that Claude Code is subscription-priced. Pro is $20/month, Max 5x is $100/month, Max 20x is $200/month, and interactive use in the terminal and IDE sits inside that subscription rather than per-token API billing. For a heavy daily user, a subscription can undercut pay-per-token on any model — and Deep​Seek Harness has no subscription; it is free software, but the tokens are pay-as-you-go. DSH's token efficiency helps close that: one hands-on test had DSH rebuild a website from a spec and chart its own GitHub growth for under three yuan (~$0.40) of total inference (DoNews, August 14, 2026), and its PTC mode and trajectory cache are explicitly designed to shrink per-task token burn.

What it is actually like to run, honestly

The setup is genuinely easy once it starts: install Node.js, run code>npx @deepseek-ai/dsh web/code>, drop in an API key, pick a workspace, pick a preset — reviewers report going from install to a finished task in minutes. The plugin ecosystem is real, not a launch-day page: 288 plugin repositories appeared on the dsh-plugin topic within 24 hours, including dsh-plan-execute (route planning to a reasoning model and execution to a cheaper one) and dsh-vision (add vision to text-only models).

Now the part the launch-day coverage skips. DSH is rough in specific, documented ways:

None of these are disqualifying for a developer-preview tinkerer. They are disqualifying for a production team.

When this recommendation is wrong

My recommendation is: run DSH for evaluation and personal projects today; do not put it in front of a production workflow yet. Here is where that advice fails, and you should ignore it:

How to try DSH today

DSH is free to install, and the whole loop takes minutes:

Because DSH accepts any Open​AI-compatible endpoint, you can point it at OrcaRouter's single API instead of managing vendor keys — DeepSeek V4 Flash 0731 and DeepSeek V4 Pro 0813 are live there at Deep​Seek's list price with zero markup, and routing V4 Flash for volume and V4 Pro for hard jobs is a config change rather than a migration. That is the setup worth testing: DSH in front, a routing layer behind it, and Claude Code kept as the fallback so a preview-version hiccup never stops your day.

Bottom line

Deep​Seek Harness is the first serious open-source challenger to Claude Code's grip on the coding-agent market — not because it is a better model (it is not a model at all), but because it changes what you own. Claude Code is a finished product you rent; DSH is a kit of parts you rebuild. Today, that means Claude Code is the right tool for most people, and DSH is the one to watch, to break, and to learn from. In three months, if the plugin ecosystem matures and the sandbox hardens, that calculus flips. Right now the decision is simple: production trust goes to Claude Code; curiosity and ownership go to Deep​Seek Harness.