What it does
7 features
DeepSeek Harness · Open-source agent harness in preview
Product and access
DeepSeek Harness is the runtime for agents, not the language model. It runs with DeepSeek models, bundled third parties or custom endpoints. CodeWhale is a separate project that, according to its repository, is not affiliated with DeepSeek.
Core features
7 features from provider sources · checked on 5 November 2026. Availability and limits are directly below.
Everything is a plugin
According to the README, the harness is built on a plugin architecture with the Cordis framework. Profiles stack plugin bundles and your own configuration layers.
Source 9Web, headless, SDK and ACP
`dsh web` starts the web interface. There are also profiles for one-off headless jobs, SDK clients and ACP. All of them use the same `dsh` command.
Source 6Work in the workspace
The agent can read and edit workspace files, run commands, delegate work and maintain a plan.
Source 10Subagents and agent teams
Subagents run as managed child sessions. The team package adds a shared task board; all members share one checkout.
Source 11Automation tasks
Reminders repeat by an RFC 5545 rule and deliver an instruction to an existing or new conversation. You can create them through the model or by hand.
Source 12Freely configurable models
Besides DeepSeek you can add bundled providers such as anthropic, openai or moonshotai. Custom endpoints speak OpenAI Chat Completions, OpenAI Responses or Anthropic Messages.
Source 13
Show all 7 features
Sandbox and permission presets
Permission presets bundle a sandbox mode with an approval policy. The local sandbox stops with an error when it is unavailable instead of running commands unconfined.
Source 14
Usable AI models
Open choice
Besides DeepSeek you choose providers from the installed catalog, such as Anthropic, OpenAI, Kimi or GLM, each with your own key. A custom API also connects gateways and self-hosted servers that speak the OpenAI or Anthropic protocols. The whole project is a developer preview, and the docs do not mention Ollama.
Model families named
- DeepSeek
- Anthropic (Claude)
- OpenAI (GPT)
- Moonshot (Kimi)
- Z.ai (GLM)
- Other models can be connected freely
- Model choice
- You choose
- Your own API key
- Yes
- Router or custom endpoint
- Yes
- Local models
- Yes
The rating covers the generally available main access. Documented · Sources accessed on
Does it fit my task?
2 practical tasks · 5 requirements and limits
A good fit if …
- you want to test an open-source harness locally and extend it with plugins
- you want to choose between DeepSeek, third parties and custom endpoints
- you can run it in a disposable VM or a container
Less suitable if …
- you need a production-ready, audited solution for sensitive data
- breaking changes between versions would be a problem for you
- you expect commitments on data location and legal framework for Europe from the provider
Editorial assessment based on provider documentation. Not a product test.
Practical tasks
Editorial examples based on documented features. No product tests performed.
Assess a repository
- Your input
- A project folder as the workspace and a request for an overview.
- Possible result
- The agent reads the code and names the main packages.
- What to check
- Start in a disposable environment and check the statements against the code.
Schedule a recurring check
- Your input
- An automation task with an instruction, a workspace and a repeat rule.
- Possible result
- The host delivers the instruction to a conversation at the scheduled time.
- What to check
- Read the result in the conversation: delivery records only confirm the delivery, not the completed work.
Have it summarize a repository and identify its main packages.
Your possible prompt
Read the project rules. Investigate [error], propose a small change, and run the relevant checks. Show me the diff and any remaining uncertainties.
Adjust the task, data and permissions.
Requirements and limits
Limitation
Developer preview without an audit
SAFETY.md calls the project experimental: no security audit, not to be treated as secure or production-ready. The README announces breaking changes. All 37 releases listed by the GitHub API are pre-releases, most recently dsh-v0.2.2-rc.1 from 4 November 2026.
Source 15Requirement
Node.js 24.9.0 or newer
The repository requires Node.js 24.9.0 or newer, and pnpm for source builds. The product page lists the desktop app only for macOS with Apple silicon and 64-bit Windows.
Source 4Requirement
Model access required
According to the terms, official models need a DeepSeek account and a balance. Otherwise you use another provider's model API with your own key.
Source 7Limitation
The sandbox is no guarantee
According to the terms, the sandbox and approvals lower the risk but do not guarantee isolation. The terms assign responsibility for agent actions to you.
Source 7Not yet confirmed
Retention of conversations
The privacy policy gives no fixed period for conversations. Network logs are kept for at least six months.
Source 8
Can I use it?
With requirements
Europe & regional availability
Read the privacy policy (as of 20 September 2026) and the terms of use. They do not mention the EU, the EEA, the UK or Switzerland. Statements on location only reflect the wording of the policy.
Reviewed scope: Official models through DeepSeek and custom models reviewed separately
Reviewed on:
Availability still unclear
Official models: storage in China
The policy says personal data collected in domestic operations is stored in the People's Republic of China and is currently not transferred abroad. Conversations may be used for training after encryption and de-identification; you can opt out through the contact channel.
Region, reason and source
Region covered: EU/EEA, United Kingdom and Switzerland
Documented reason: The policy makes no separate statement for users in Europe; availability and legal bases there remain open.
With requirements
Custom models: data with the chosen provider
With custom models, according to the policy, DeepSeek Harness does not store or process inputs and outputs on its servers. They go to the provider you configured, whose terms apply.
Region, reason and source
Region covered: EU/EEA, United Kingdom and Switzerland
Documented reason: The data path depends on the configured provider and on plugins or MCP servers that process data themselves.
Availability still unclear
Legal framework of the terms of use
The terms place the contract under the law of mainland China, with the courts at DeepSeek's domicile. They name no regional restrictions.
Region, reason and source
Region covered: EU/EEA, United Kingdom and Switzerland
Documented reason: The reviewed pages do not say whether different European rules apply.
Accounts & permissions
Based on provider documentation. Permissions depend on the account, integration and organization policies. Review the permissions actually requested before use.
Configuration covered: Locally run harness; an account only for official models
Sources reviewed:
Show all 3 points
Documented with conditions
Data access & team sharing
According to SAFETY.md, the project can run code and commands, load third-party plugins and access the network, processes, credentials and files made available to it. How far this reaches depends on the sandbox and the environment.
Evidence
Memory and conversation continuity
0 of 6 criteria documented
The shipped composition includes no memory server. Three MCP examples are off by default; storage and retrieval depend on the chosen system.
No built-in memory; optional MCP servers
Sources accessed:
Documentation by criterion
StoragePartly documented
Storage and scope
The shipped composition has no memory server. The examples store data where the third-party system decides, such as `~/.memorix/data`, `~/.engram` or a JSONL file in the home directory.
SelectionPartly documented
Selection and timing
DSH makes no selection of its own. The docs optionally suggest an instruction: on request the model writes through a memory tool. Selection rules belong to the third-party system.
Edit & deleteNot publicly established
Inspect, edit and delete
The page describes no interface of its own for review, editing or deletion. This depends on the chosen server.
Access boundariesPartly documented
Agent and channel boundaries
According to the page, Memorix and Engram derive the project from the Git project in the DSH working directory. The page does not describe boundaries between agents or channels.
RetrievalPartly documented
Retrieval
Retrieval runs through the server's MCP tools. The reference server searches by substring, not semantically.
New conversationsPartly documented
Session reset and compaction
Compaction summarizes older history and also runs through `/compact`; the original stays in the session log. The package docs describe no link to memory.
Cost and effort
3 plans · Prices as of 5 November 2026
Prices and included usage
Pricing checked
The product page names no price. The software is MIT licensed; costs arise with the model provider. DeepSeek publishes an API price list.
| Plan | Price | Included |
|---|---|---|
| DeepSeek Harness | Free under the MIT license | Software with web interface and plugins; model use not included. Price source |
| deepseek-flash (API) | Input without cache $0.15 to $0.3, output $0.6 to $1.2 per 1M tokens (off-peak to peak) | Model version DeepSeek-V4.1-Flash-1017; cache hits $0.003 to $0.006. Price source |
| deepseek-v4-pro (API) | Input without cache $0.66 to $1.32, output $1.98 to $3.96 per 1M tokens (off-peak to peak) | Model version DeepSeek-V4-Pro-0813; cache hits $0.022 to $0.044. Price source |
- The API price list uses the $ sign without a currency code. The reviewed pages do not say whether official models on a Harness balance are billed exactly this way.
- Third-party pricing was not used as evidence.
Setup
Install Node.js 24.9.0 or newer and run `npx @deepseek-ai/dsh web`, or download the desktop app, or build from source with pnpm. Then add model access and choose a workspace. The project advises using a disposable VM, a container or a dedicated environment.
Operation & oversight
Based on public provider documentation. Assessments are editorial inferences from the stated properties. Actual time requirements and reliability have not been measured.
Deployment covered: Self-run process on your computer or server
Sources reviewed:
Editorial assessment
The harness describes the sandbox, approvals and sessions at length in the repository. It remains a preview without an audit, though. You run it yourself and, according to the terms, bear the consequences of its actions.
Show all 6 points
Documented with conditions
Resuming after failures
Saved sessions can be resumed. After a context overflow, the basic package condenses and retries. After a restart, the host sends an overdue one-shot task or the latest missed occurrence; exactly-once delivery is not guaranteed.
Evidence · Evidence · Evidence
Evidence and open questions
24 sources · 3 open points
Open points
The reviewed sources contain no reliable information on these points. A gap says nothing about the product.
Europe: Availability still unclear
Requirements and limits: not yet confirmed
Sources and evidence
Read the repository (README, SAFETY.md, docs, package READMEs, release list via the GitHub API), the product page, the privacy policy, the terms of use and the API price list. Feature claims come from the provider; example tasks are editorial. Our own hands-on tests are pending.
Editorial team: Michael Meierhoff · sources reviewed on . Provider information is not independent evidence of performance.
Provider
According to the terms and privacy policy, the company behind the product is Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. It also develops the DeepSeek language models.
- Operator
- DeepSeek
- Organization
- Provider
- Website
- deepseek.com
Changes to this profile 1
Open-source agent harness from DeepSeek with a plugin architecture, web interface and desktop app, currently a developer preview.
These notes describe editorial changes. Source and pricing review dates appear in their respective sections.