DeepSeek Harness is DeepSeek’s August 2026 developer-preview Agent runtime, open-sourced under MIT at deepseek-ai/deepseek-harness. It does not ship “another chatbot model.” It answers: once a model can think, how do you reliably split work, call tools, edit files, run commands, log sessions, and recover—summarized as Model + Harness = Agent.
This site documents Clash Verge Rev. If you run Harness on a restricted network, shells often ignore system proxy even when the browser looks fine. Concepts here; wiring: Harness + Clash Verge. Sibling agents: Claude Code, Codex CLI.
One sentence
If the model is the brain, the harness is the hands and the journal: it decides what to call, runs it inside a boundary, records what happened, and picks up where it left off. You debug the harness when tools fail, not when the model is merely “thinking.”
Split model weights from the harness runtime
When headlines say DeepSeek open-sourced something, ask which layer. Weights set reasoning quality and price. The harness decides whether the model can finish real repo work. Without it you mostly chat; with it the model operates inside a controlled tool boundary and leaves an auditable session log.
| Layer | Typical artifact | What you debug |
|---|---|---|
| Model | API / weights | Keys, quotas, latency, policy |
| Harness | Framework, plugins, Web UI, loop | Tool rights, resume, plugin clashes, local net |
| Proxy client | Clash Verge, etc. | Whether the shell is tunneled, region, DNS |
Mixing “API timeout” with “plugin missing” into one “Harness is broken” sentence burns an afternoon on the wrong layer.
Everything is a plugin
Public materials describe a pluginized kernel (often Cordis): model adapters, tool packs, session storage, UI, and even the agent loop load as plugins. The point is to extend via config and plugins instead of forever forking the trunk loop.
Practically you can swap backends, shrink the tool surface (minimal modes for evals), or experiment with creative plugin mixes. Mode names move in a preview—treat README as source of truth, not a blog checklist as a contract.
How to try it (follow upstream)
You typically need a current Node.js toolchain. The documented fast path is npx for the Web UI (step-by-step: npx one-command; full tree: git clone install):
- Install Node.js (version per README).
- Run
npx @deepseek-ai/dsh web(if the package name changes, follow the repo). - Configure a DeepSeek (or other adapted) API key.
- Accept with a minimal task—list a directory or read one small file—not “rewrite the monorepo.”
Or clone https://github.com/deepseek-ai/deepseek-harness for plugin work. MIT permits commercial use; you still own preview stability and compliance reviews.
Versus Claude Code and Codex
- Claude Code / Codex: packaged entrypoints with relatively fixed debug paths (see our guides).
- DeepSeek Harness: composable runtime, DeepSeek-friendly by default, adapters for more; you own plugins and config.
Need one script fixed today? A packaged CLI is usually faster. Need your own tool allowlists, audit logs, and multi-model swaps? A harness fits better. Skip “which is stronger”—ask whether you need to own the runtime.
When Clash Verge enters the picture
DeepSeek’s API is often reachable without a tunnel. Agents that npm install, hit GitHub, or fetch overseas docs fail while the browser still works. Overview: Tun & AI. Step-by-step: Harness proxy guide. Builds: download center.
Acceptance and stop-loss
Concept: you can separate model failures from harness/tool failures. Trial: UI up, key set, minimal read/list succeeds. Network: outbound tasks leave rows on Clash connections. Backup sessions/plugins before preview upgrades. Change one layer at a time when stuck.