Rating certificate · AW/26-09/57826 Sept 2026
66of 100
DeepSeek Harness (dsh)
DeepSeek AI · Coding agents
No. 9 of 16 in categorymedium confidenceTrust BOpen source
The Register / Coding agents / DeepSeek Harness (dsh)
Register entry · Coding agents · updated 26 Sept 2026
DeepSeek Harness (dsh)
DeepSeek Harness (dsh) is an MIT-licensed agent harness in which models, tools, sessions, sandboxes, storage, the agent loop and the UI are all replaceable plugins, built on the Cordis plugin kernel. It is launched with `npx @deepseek-ai/dsh web` (local web UI) and also offers headless, ACP and SDK profiles and a preview desktop app. It defaults to the DeepSeek API and supports third-party providers (Anthropic, OpenAI, Moonshot, Z.ai) and custom OpenAI- or Anthropic-compatible endpoints including local servers.
In beta. Public since 2026-08-13 and labelled a developer preview: README states 'THERE WILL BE COMPATIBILITY-BREAKING CHANGES' and SAFETY.md says it 'has not undergone a security audit and must not be treated as secure or production-ready'. All 22 GitHub releases in the last 90 days are prereleases (alpha/rc); latest dsh-v0.1.7-rc.2 on 2026-09-24.
What reviewers and users praise
- Highly composable architecture (Wavect, Flowtivity and DeepInfra reviews)
- Fast and cheap with DeepSeek API or local models; high cache hit rates reported (HN users)
- Heavy early adoption and community plugin activity (7,790 GitHub Discussions in six weeks)
What they criticise
- Developer preview with breaking changes; not security-audited (README, SAFETY.md)
- Sandbox confines writes only; reads, network and (on some backends) process visibility are not confined (Wavect, DeepInfra, CLI reference)
- Full session logs (messages, tool arguments/results, workspace paths) uploaded to DeepSeek by default when using the DeepSeek API, since 2026-09-14 (repo decision note)
- Preview-stage bugs: data loss on upgrade, runaway token loops, Windows sandbox failures (GitHub Discussions)
Experts vs the crowdcrowd warmer by 9
Experts50
Crowd59
| Price | Free |
|---|---|
| Use it via | Web app, CLI, Desktop app, API |
| Runs on | macos, linux, windows |
| Licence | MIT |
| Updated | Evidence collected 26 Sept 2026 |
How the rating was worked out
Weights for Coding agents: adoption 35%, experts 40%, crowd 25%. Pillars without enough evidence are left out and the others re-weighted. The method.
| Adoption & momentumweight 35% | 90 | GitHub stars 236k → 100 · npm weekly downloads 473k → 72 · 22 releases in 90 days → 100reach × 0.7 + momentum × 0.3 |
|---|---|---|
| Expertsindependent reviews · weight 40% | 50 | 3 independent reviews (0 positive · 3 mixed · 0 negative), averaged toward 50 for small samples. |
| Crowdcommunity sample · weight 25% | 59 | 11 dated posts from hackernews, github: 6 positive · 2 mixed · 3 negative (positive + ½ mixed) ÷ n, averaged toward 50 for small samples |
| Rating | 66 | medium confidence (3 of 3 pillars, 11 community posts). |
Trust & safety
Seven dimensions graded A–D from documented facts. Overall: B.
A
Permission model
New sessions default to the workspace-write preset: Bash and file mutations confined to the workspace via bwrap/Landlock/Seatbelt/Windows ACL; UI asks before operations requiring approval under the active policy. Caveats: web_fetch runs without per-call confirmation in every mode; the sdk-minimal profile pins danger-full-access; SAFETY.md says sandboxing does not guarantee isolation. source
B
Data access scope
Writes scoped to the session workspace, but 'reads and network access are not confined'. Broad read access, documented. source
A
Data storage
Self-hosted; sessions and credentials stored locally ($DSH_HOME). Note: session-log upload to the DeepSeek API endpoint is on by default but can be disabled (enabled: false), so cloud storage is not required. source
n/e
Data retention & training
Since 2026-09-14 the default sends the canonical session log (message text, tool arguments and results, workspace paths, feedback) to the DeepSeek endpoint with each eligible request. DeepSeek's training/retention terms for this data were not checked (a 'Data Processing Statement' is linked from the product page but not read). source
A
Incident history (24 months)
No published GitHub Security Advisories and none in the GitHub Advisory Database for @deepseek-ai/dsh; project is six weeks old and GitHub Issues are disabled (reports go to Discussions), which limits visibility. Security-relevant bug reports exist in Discussions (e.g. #7904 auto-review deny bypass). source
D
Compliance
SAFETY.md: 'It has not undergone a security audit'; no attestations. source
A
Transparency
MIT source, public releases, extensive in-repo architecture and decision notes. source
The experts
3 named independent reviews found; dated reviews from the last 12 months are scored.
- mixed
Use it when replaceable agent infrastructure is a strategic requirement and your team is prepared to own the composition.
Kevin Riedl (Wavect) · 16 Aug 2026 · wavect.io - mixed
this is a v0.1 developer preview with explicit breaking-change warnings
AJ Awan (Flowtivity) · 20 Aug 2026 · flowtivity.ai - mixed
A rambling idea note and two prompts produced a tool that works
Stefan Fidanov (DeepInfra blog) · 9 Sept 2026 · deepinfra.com
The crowd
11 coded posts from hackernews, github.
6 positive2 mixed3 negative
- mixed
I'm using the Deepseek harness with the deepseek computer use and deepseek browser integration, it's somewhat janky but it gets the job done.
news.ycombinator.com · 25 Sept 2026 · news.ycombinator.com - mixed
Used to be I preferred DSH, but zcode has some features I like over DSH.
news.ycombinator.com · 23 Sept 2026 · news.ycombinator.com - positive
opencode pushes CPU usage to very high which has NOT happened with DeepSeek harness even once.
news.ycombinator.com · 22 Sept 2026 · news.ycombinator.com - positive
DeepSeek Harness is my favorite for coding.
news.ycombinator.com · 18 Sept 2026 · news.ycombinator.com - positive
I just got DeepSeek Harness (DSH) set up with 2x R9700 and it's rather mind blowing that these can do actual work and quickly.
news.ycombinator.com · 8 Sept 2026 · news.ycombinator.com - positive
now with the deepseek harness, I don't even miss Claude Code.
news.ycombinator.com · 1 Sept 2026 · news.ycombinator.com - positive
I've ran 1 billion tokens through with >3000 steps and multi-hour autonomous output with nothing more than a 5090, Deepseek Harness, and Qwen 3.8 27b
news.ycombinator.com · 31 Aug 2026 · news.ycombinator.com - positive
I did start using the Deepseek harness and v4 flash due to the crazy low usage limits on sol and I'm quite pleased
news.ycombinator.com · 22 Aug 2026 · news.ycombinator.com
Pricing
As published, checked 26 Sept 2026.
| Plan | Price | Notes |
|---|---|---|
| DeepSeek Harness software | Free | MIT; no subscription found on the product page. |
| DeepSeek API (pay-as-you-go) | Custom | Default model provider; per-token pricing not collected. |
Recent changes
Ranked alongside
Other ranked entries in Coding agents.
OpenCodeAnomaly75
Claude CodeAnthropic73
Zed (Agent Panel)Zed Industries73
GitHub CopilotGitHub (Microsoft)72
Codex (CLI and app)OpenAI71
Sources
Every figure above traces to one of these, observed up to 26 Sept 2026.