DeepSeek Harness is an open-source agent runtime built around Cordis: tools, sessions, storage, scheduling, and the web UI are replaceable plugins. Install Node.js, run npx @deepseek-ai/dsh web, then configure your API key and workspace. Start with Standard mode. Use Create when you are ready to experiment with plugins. The harness is free to install; model API usage is billed separately.
What is DeepSeek Harness?
DeepSeek puts it as a formula: Agent = Model + Harness. LangChain used the same framing earlier when writing about agent harnesses, and it holds up. The model does the thinking. The harness is everything around it: tools, a skills system, sessions, a sandbox, storage, the agent loop, scheduling, sub-agents, workflows. Claude Code and Codex are harnesses in this sense, and a harness with no model attached is an empty shell.
The supplied write-up describes the August 2026 developer preview, following V4 Pro’s stable release. It reports 37,000 GitHub stars in roughly a day, more than 100 first-party plugins (160 in its settings screenshot), and 577 community repositories. These are observations from that source, not live counts.
The original article contrasts Harness with packaged coding assistants. Its focus is the depth of customization: beyond skills and MCP servers, Harness exposes the components that run the agent as replaceable plugins. The available extension points in other products vary by product and version.
DeepSeek Harness goes the other way. Every one of those pieces is a plugin: tools, the sandbox, storage, scheduling, sub-agent orchestration. Even the web interface you type into is a plugin, which is why skin plugins for it already exist. The comparison the original write-up reaches for is the modular phone: every component is owner-replaceable.
Why is it called "Harness" and not "DeepSeek Code"?
Because the product is not the point. DeepSeek shipped this as a developer preview and says on its own site that it is iterating fast, with breaking changes to come. The name matches the intent. "Code" would have framed it as another coding assistant. "Harness" frames it as infrastructure that other people build on. The bet is that developers will write plugins for it, grow an ecosystem around it, and push agents toward something the write-up calls self-evolution: agents that extend themselves.
This is also why the release split the community so hard. One camp calls the plugin-first design the future. The other asks why anyone would hot-swap components mid-task when a normal product just works. Both readings are defensible, and the fastest way to form an opinion is to install it.
What does the Cordis kernel actually do?
The kernel at the center is called Cordis, and it is deliberately small. It loads plugins, unloads plugins, and manages dependencies between them. That is the whole job description.
The value sits in the guarantees behind those two operations. There is an 88-page paper behind the design, "A Programming Paradigm for Spatiotemporal Composability" by Yifan Shu, Wei Zhang, and Tianyi Cui (Peking University and DeepSeek-AI). Two properties carry it. Temporal composability: unloading a plugin fully reverts the side effects it produced, so nothing it did lingers in the context. Spatial composability: when the plugins a component depends on appear, disappear, or change, it re-derives its dependencies instead of breaking.
Together they make hot-swapping safe enough to do while the agent is running, and that is the design's real point: an agent that can load and unload its own capabilities at runtime can extend itself. The write-up argues Harness is closer to a research artifact than an agent product, a Cordis showcase that ships with more than 100 first-party plugins attached. The preview build's settings panel lists 160.
How to install DeepSeek Harness and start your first chat
Install Node.js first. Then open a terminal and run:
npx @deepseek-ai/dsh web- npm pulls the package (the build shown in the source ran dsh@0.1.0-rc.6) and prints a local URL. In that run it was
http://127.0.0.1:3080. Open it in a browser. If the command itself is the hurdle, paste it into any local agent you already use and let it do the install. - First launch asks for a DeepSeek API key. Get one at platform.deepseek.com and put some balance behind it, because the welcome screen does not move without it. After that you land on the home screen. Pick a workspace, which is the project directory bounding what the agent can touch, and start a chat.
- The model picker has two entries,
deepseek-v4-flashanddeepseek-v4-pro, each with a reasoning effort setting: Off, High, or Max.
Version context: the steps describe the source’s 0.1.0-rc.6 preview. The official pricing page now lists deepseek-flash for DeepSeek-V4.1-Flash-1017; deepseek-v4-flash is a legacy alias. The install command is documented by DeepSeek; installation has not been independently tested for this guide.
What does DeepSeek Harness cost?
Current official API snapshot
Checked October 19, 2026. USD per million tokens. Flash uses deepseek-flash (V4.1-Flash-1017); Pro uses deepseek-v4-pro (V4-Pro-0813).
Official API rates · USD / 1M tokens
| Model / period | Cached input | Uncached input | Output |
|---|---|---|---|
| Flash · peak | $0.006 | $0.30 | $1.20 |
| Flash · off-peak | $0.003 | $0.15 | $0.60 |
| Pro · peak | $0.044 | $1.32 | $3.96 |
| Pro · off-peak | $0.022 | $0.66 | $1.98 |
Peak: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays. All other hours, including weekends and Chinese public holidays, are off-peak. Provider prices can change.
Historical context: the source’s August 2026 tables
The supplied article records a pricing change effective at 00:00 Beijing time on August 17, 2026, with peak windows of 09:00–12:00 and 14:00–18:00 Beijing time. Its original CNY tables and chart are retained below as historical context. Use the current official snapshot above when checking the present model names and billing schedule.
Historical DeepSeek pricing · CNY / 1M tokens
| Model | Billing item | List price through Aug 16 | Peak from Aug 17 | Off-peak |
|---|---|---|---|---|
| deepseek-v4-flash | Cache-hit input | ¥0.02 | ¥0.10 (5×) | ¥0.05 |
| deepseek-v4-flash | Cache-miss input | ¥1.00 | ¥3.00 (3×) | ¥1.50 |
| deepseek-v4-flash | Output | ¥2.00 | ¥9.00 (4.5×) | ¥4.50 |
| deepseek-v4-pro | Cache-hit input | ¥0.025 | ¥0.30 (12×) | ¥0.15 |
| deepseek-v4-pro | Cache-miss input | ¥3.00 | ¥9.00 (3×) | ¥4.50 |
| deepseek-v4-pro | Output | ¥6.00 | ¥27.00 (4.5×) | ¥13.50 |
In the historical table, V4 Pro’s peak cache-hit input is 12 times the earlier listed rate. The source compares its ¥9 cache-miss input and ¥27 output with GLM-5.2’s ¥8 and ¥28, and Kimi K3’s ¥100 output. These are comparisons within that snapshot, not current quotes.
Historical cross-provider comparison · CNY / 1M tokens
| Model | Cached input | Uncached input | Output |
|---|---|---|---|
| GLM-5.2 | ¥2.00 | ¥8.00 | ¥28.00 |
| Kimi K3 | ¥2.00 | ¥20.00 | ¥100.00 |
| Qwen 3.8 Max | not listed | ¥12.00 | ¥36.00 |
| GPT-5.6-Sol (input up to 272k) | ¥3.39 | ¥33.94 | ¥203.66 |
| Claude Fable 5 | ¥6.79 | ¥67.89 | ¥339.44 |
| DeepSeek V4 Pro, peak (from Aug 17) | ¥0.30 | ¥9.00 | ¥27.00 |
| DeepSeek V4 Pro, off-peak | ¥0.15 | ¥4.50 | ¥13.50 |
Historical source snapshot, compiled August 14, 2026, with DeepSeek’s August 17 rates added. All figures are CNY per million tokens. These original figures have not been independently verified against archived provider price lists.
In the source’s pricing example, running the same token workload off-peak halves the model charge compared with peak hours. Harness also lets you configure additional providers from its catalog, or custom providers with your own base URL, protocol, and model list. The original article demonstrates GLM inside Harness.
Planning to use a coding agent for ongoing work? Explore SaveMyToken’s Coding Plan.
What are the four modes?
The mode picker confuses people, so the framing first: these four are template presets that decide which plugins load. Underneath, it is the same Cordis kernel either way.
1. Standard
The full coding agent with everything prewired: file reading and editing, shell, file search, web search, skills, plans, goals, background tasks, sub-agents, workflows. If you are new, or if you just want to get work done, use this and do not look back.
2. PTC
Everything Standard has, plus a Code Mode SDK. Instead of one tool call at a time, the model writes a TypeScript program that batches several tool operations into a single run_code execution. Five round trips of read, search, filter, and merge collapse into one program run. It saves tokens and fits structured, multi-step, parallel work. It also needs a model with stable code planning, and debugging a misbehaving program is harder than debugging five tool calls, so it is not a beginner setting.
3. Minimal
Two tools: a persistent bash and a file editor. The system prompt is pinned to one line, "You are a helpful software engineering assistant," and context compression is off. This exists to benchmark bare models. It is the right tool for comparing two models' raw agent ability and the wrong tool for everything else.
4. Create
The one that shows why the architecture exists. You get Standard's full ability set, plus the right to inspect the running Cordis environment, trial plugins in memory, and build new agents and plugins, including attaching a freshly built plugin to its own live pipeline. Ask it for "a mode that can read code but never modify files, for security audits," or "a research agent wired to internal search with three custom skills," and it constructs them at runtime. The write-up's image: the agent notices it has no wrench, forges one on the spot, bolts it onto itself, and gets back to work.
Why does the append-only session log matter?
A Harness session is an append-only event log. System prompts, user messages, reasoning content, tool calls and results, permission changes, context injections, compaction events, sub-agent dispatches: all of it lands in the log as events. The history the model sees on each turn is re-derived from that log instead of accumulating in place.
The payoff is the trace view, which lets you inspect any run by source. When an agent fails, the usual experience is a dead task or an infinite loop with no idea where it went wrong. Here you can find the step where it went off the path. That makes runs observable, auditable, and reproducible, and for anyone doing serious agent work it is not a nicety.
Which community plugins are worth installing first?
The dsh-plugin topic on GitHub counted 577 public repositories when the source looked in August 2026. From that pile, five picks:
Community examples from the source. These packages have not been installed or reviewed for this guide; compatibility depends on your Harness version.
- dsh-at-file — Type @ in the input box to pull any file into context.
- dsh-genui — Renders charts, tables, forms, diffs, Mermaid diagrams, and interactive panels directly in replies.
- dsh-automation — Adds scheduled and recurring runs. A real gap in the core product, though the interaction is rough at the edges.
- DSH-better-sidebar — A VS Code-style workbench in the sidebar: file management, code editing, a real terminal, git, diffs, an embedded browser, background tasks, and sub-agent status.
- ModLens — Gives text-only DeepSeek models a vision channel. Paste an image into the conversation once it is configured.
Should you use it?
Split verdict. The plugin architecture is real rather than marketing: the two composability guarantees are written up in an 88-page paper and implemented in the kernel. The product around them is rough by the standards of consumer software. The vocabulary assumes you are a developer. Onboarding costs more effort than it should, the feature set is thin in places, and the UX asks for patience. Any one of those sheds casual users. The write-up's read is that DeepSeek knows and does not mind: it is a research-first team shipping infrastructure in public, under a tagline that translates roughly as "explore the realm not yet reached."
If you want a familiar setup, start with the coding agent you already use. If you want to experiment with agent infrastructure, especially a kernel that lets an agent extend its capabilities during a run, Harness gives you a concrete system to explore.
FAQ
What is DeepSeek Harness in one sentence?
An open-source agent harness from DeepSeek, built on the Cordis kernel, in which tools, skills, sessions, the sandbox, storage, scheduling, and even the web UI are hot-swappable plugins.
Is DeepSeek Harness free?
The harness is open source and free to install. You pay for model API usage, starting with a DeepSeek API key, or with any compatible provider you configure yourself.
Does it lock me to DeepSeek models?
No. Model settings accept catalog providers and custom providers with your own base URL, protocol, and model list. The source demonstrates running GLM inside Harness.
How is it different from Claude Code or Codex?
Harness makes its runtime layers available as plugins you can replace or write yourself. Other coding agents expose different extension points; the exact comparison depends on the product and version. The source article describes Harness as the rougher experience.
Which mode should a beginner choose?
Standard. PTC requires stable code planning from the model and is harder to debug, Minimal exists for benchmarking, and Create mode makes the most sense once you understand the plugin system.
How much did DeepSeek API prices go up?
The source’s August 2026 snapshot shows increases of 3× to 5× on most items and 12× for V4 Pro’s peak cache-hit input. Those are historical comparisons. Check the current official snapshot in the pricing section before estimating a new workload.