DeepSeek Harness
Open-source agent runtime from DeepSeek. Every piece — models, tools, sandboxes, UI — is a plugin. MIT-licensed developer preview launched August 13, 2026.
Updated 2026-08-17
Yes. DeepSeek Harness is free to use.
Listed pricing: Free (MIT). Inference is whatever model you plug in — DeepSeek V4-Flash and V4-Pro are the vendor pairing.
Overview
DeepSeek Harness is a developer-preview agent runtime, announced August 13, 2026 alongside DeepSeek-V4-Pro. The official page and the GitHub repo use the same slogan: "Everything is a plugin." Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are described as swappable pieces on a Cordis kernel. License is MIT. The public repo is github.com/deepseek-ai/deepseek-harness.
Quick start from the vendor: npx @deepseek-ai/dsh web for a local UI, or clone and use pnpm. Modes listed in the preview include Standard, Code, Minimal, and Creator. Every run is supposed to leave an append-only event stream; Trajectory view is the replay and fork surface. V4-Pro launch coverage notes that some of the published Pro agent scores were run in Minimal mode — that is a vendor/press claim, not an independent bake-off.
This is infrastructure, not a hosted assistant. You still pay for whatever model you point it at. DeepSeek's own pairing is V4-Flash (high concurrency, lower price) and V4-Pro (harder agent tasks). It is the open-runtime counterpart to Claude Code or Codex, not a replacement for the DeepSeek API.
Keep the repo and deepseek.com/harness as the live pages. Star counts move daily; do not treat a screenshot of GitHub as a durable fact. Confirm the latest tag and the npx package name on the README before you paste a command into a tutorial.
Is DeepSeek Harness free?
Yes. DeepSeek Harness is free to use.
What the free tier covers: The runtime is free. Model tokens are billed by whichever API you configure.
Listed pricing: Free (MIT). Inference is whatever model you plug in — DeepSeek V4-Flash and V4-Pro are the vendor pairing.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
DeepSeek Harness pricing
| Plan | Price | What's included |
|---|---|---|
| Open source | $0 | MIT license, source on GitHub. You bring model keys. Developer preview (v0.1) as of August 13, 2026. |
MIT license, source on GitHub. You bring model keys. Developer preview (v0.1) as of August 13, 2026.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is DeepSeek Harness worth it?
Worth it for Developers who want an MIT-licensed agent runtime they can recompose, and who will bring their own model keys.
You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8.4/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with DeepSeek Harness. How we verify.
Worth it if
The strengths recorded against this entry.
- MIT-licensed; you can read and fork the loop
- Plugin model means you are not stuck on DeepSeek weights
- Launched next to V4-Pro, so the vendor's own agent scores have a named runtime
- Local npx path, no hosted lock-in for the runtime itself
Not worth it if
Any one of these blocks your use case.
- Developer preview v0.1 as of August 13, 2026 — expect breakage
- You still pay for inference, and you assemble the stack
- Not a product with an SLA or a consumer UI beyond the local web mode
- "Rival to Claude Code" is press framing, not a documented feature-parity claim
What sets DeepSeek Harness apart
- Every capability is a plugin on the Cordis kernel
- Append-only session logs with a Trajectory view for replay and fork
- Standard, Code, Minimal, and Creator modes in the preview
- Local npx start; no hosted lock-in for the runtime itself
Key features
Plugin architecture
Vendor description: every component (models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, UI) is a plugin on the Cordis meta-framework.
Traceable sessions
Append-only session logs. Trajectory view is documented for replay and fork.
Modes
Preview lists Standard, Code, Minimal, and Creator. Press notes that some V4-Pro agent benchmarks used Minimal mode.
Local quick start
npx @deepseek-ai/dsh web, or git clone plus pnpm, per the Aug 13 preview page.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| DeepSeek Harness | Developers who want an MIT-licensed agent runtime they can recompose, and who will bring their own model keys. | Free (MIT). Inference is whatever model you plug in — DeepSeek V4-Flash and V4-Pro are the vendor pairing. | 8.4/10 |
| DeepSeek | Developers and teams running heavy coding or reasoning workloads who want frontier-level model performance without frontier-level API bills. | Free chat + API from $0.27/M input tokens | 8.9/10 |
| Claude Code vs Claude Code → | Developers who want an agent that works inside an existing repo and toolchain rather than in a hosted editor, and who already pay for a Claude plan. | Included with Claude Pro $17/mo, Max 5x $100/mo, Max 20x $200/mo, Team and Enterprise plans | 9.3/10 |
| OpenAI Codex | Teams already standardised on ChatGPT who want the coding agent bundled with the seats they pay for, and who want the same agent in a terminal, an editor and a pull request. | Included with ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise plans | 8.8/10 |
| Grok Build | Developers who want a terminal agent on Grok 4.6 with a plan you approve before edits. | Product page says you can try it free; SuperGrok and X Premium+ still listed as paid paths | 8.2/10 |
Compare head-to-head
Related reading
DeepSeek Harness: Open Agent Runtime, How to Run It
DeepSeek open-sourced an MIT-licensed agent runtime on August 13. Plugin architecture, Trajectory replay, and the npx command to run it locally.
GPT-6 Sol and Luna: half the price, mixed scores
OpenAI's GPT-6 Sol and Luna cut token prices roughly in half and stay near GPT-5.6 on Artificial Analysis indexes, with lower hallucination and weaker
GPT-Live's one stream vs Gemini 3.8 Live
OpenAI's GPT-Live listens and speaks on one WebSocket. Google's Gemini 3.8 Live reasons while talking. Specs, prices, and who each model is for.
Ready to try DeepSeek Harness?
Head to the official site to start with DeepSeek Harness — pricing and plans are listed above.
Visit DeepSeek Harness