Blog · Analysis
DeepSeek Harness vs America's agent tools — the numbers that matter
August 18, 2026 · dshbase
On August 13, DeepSeek published an open-source agent harness. Five days later it had ~158,000 GitHub stars — more than OpenAI's Codex (~106k) and within shouting distance of Anthropic's Claude Code (~141k). Nobody in the "AI coding tools" business expected the agent-harness category to produce a rocket this fast. This isn't a review of whether DSH is good. It's a look at what that trajectory does to the business model of America's closed coding agents — and what we, as a plugin directory that actually installed 1,777 of these plugins, are seeing from the ground.
The star data, straight from GitHub
We pulled live numbers so this isn't vibes:
| Tool | GitHub stars | License | Model lock-in |
|---|---|---|---|
| DeepSeek Harness (5 days old) | ~158,000 | MIT, open | Any OpenAI-compatible (DeepSeek + others) |
| Claude Code | ~141,000 | Closed | Anthropic models |
| OpenAI Codex | ~106,000 | Closed | OpenAI models |
| Cursor | Proprietary app | Closed | Bring-your-own, defaults to frontier |
| GitHub Copilot | Proprietary | Closed | Microsoft/OpenAI stack |
The honest caveat: stars aren't revenue, and open-source adoption isn't a business. But the velocity is the signal. A Chinese open-source project overtaking Codex's star count in under a week reframes the whole category as "harness + plugins" rather than "vendor app + subscription".
Why the harness is eating the app
Claude Code and Codex are products: a vendor picks the model, the UX, and the extension mechanism, and you rent the whole thing. DSH is a harness: a shell that assembles plugins — model, tools, UI, even the agent loop itself — into an agent you own. The difference sounds abstract until you see what it unlocks:
- The model isn't the moat. Point DSH at DeepSeek, an Anthropic key, or any OpenAI-compatible provider. The harness follows your cost and compliance needs; a product doesn't.
- Plugins are the moat — and they're open. npm is the distribution layer. Anyone ships one; no app store review, no vendor gate.
- The agent is a file you can version. Presets and sessions are append-only logs you can rebuild, diff, and share. That's not a feature of the closed apps; it's a category difference.
We've been living this for five days at dshbase: 1,777 plugins cataloged, 1,580 verified by actually installing them on dsh 0.1.0-rc.6, 315 installable from npm with one command. When an open ecosystem produces a directory that big in under a week, the closed apps' curated marketplaces have to answer a harder question than "which one has better autocomplete."
What actually gets pressured
Pricing, and specifically the subscription
Claude Code sits behind a Pro/Max subscription; Codex behind OpenAI credits; Cursor at $20–200/month. DSH is free and self-hostable. The pressure isn't that everyone will switch tomorrow — it's that the reference price for "a coding agent" just dropped to $0. Enterprise buyers who were already questioning per-seat pricing now have a credible "build it on the harness" alternative to negotiate against.
The extension ecosystem, which is the real moat
The closed tools bet on curated Skills/MCP marketplaces. DSH bet on everything-is-a-plugin over npm. When the community ships 1,777 plugins and the harness lets you install any of them in one command, the moat moves from "what the vendor curated" to "what the community built" — and that's not something a closed app can match without opening up its own extension model.
The cost-race default
Across our cost analysis, harness choice moves the token bill by up to 7x on the same model. DeepSeek's pricing + DSH's append-only session design means the cheap option isn't a workaround anymore; it's the default. American tools tuned for their own models simply don't capture DeepSeek's cache-friendly economics.
What does NOT get pressured (yet)
Let's be honest about the limits:
- Polished UX. Cursor's editor integration and Copilot's IDE ubiquity are real products. DSH's "everything is a plugin" is powerful and also a lot of assembly.
- The plugin ecosystem is young. We verified 1,777 plugins, but the failures are real — installs break, some need a web runtime, quality varies. That's the price of openness.
- Enterprise support. Closed vendors sell SLAs, compliance, and hand-holding. Open source sells itself. Different buyers.
FAQ
Is DeepSeek Harness a threat to Claude Code?
Not as a drop-in replacement today — but as a force that resets the category's economics, yes. Star velocity aside, it changes the reference price and the extension model. The closed tools will have to respond on openness or price.
Which of the American tools is most at risk?
The one whose moat is just the model or just the subscription. If a tool's differentiation is "we tuned our model well," DSH's model-agnostic harness undercuts it. The safest are the ones with deep editor/IDE integration and enterprise distribution.
Does open source actually win here?
Not automatically. The closed apps still have better onboarding and support. What's changed is that "open" now has a credible, fast-moving reference implementation — and a plugin directory with 1,777 verified entries that any buyer can check before choosing a vendor.
Check the verified data yourself: browse all 1,777 plugins, read our 12-tool comparison, or dive into why harness choice moves cost 7x.