Wink Pings

DeepSeek Harness Entering the Enterprise Gets Stuck on Permissions and Auditing: Two Open-Source Projects Offer Different Solutions

dsh-ee releases v0.12, bringing connectors, skills, and experts into employees' local DSH; OwnDsh builds a self-hosted control plane to manage identities, model keys, quotas, and auditing. The two take different routes, but both aim to let enterprises connect internal systems to AI.

DeepSeek Harness lets employees run AI inference and tool calls on their own computers. For enterprises that want to use it, the tricky parts are permissions, data, and auditing. Two recent open-source projects are filling this gap, in different ways.

dsh-ee: Assembling Company Systems and Knowledge into Micro-Apps Employees Can Use

The author of dsh-ee introduced the project in a tweet. It is positioned as an enterprise-grade deepseek-harness solution, fully open source under Apache 2.0. v0.12 officially supports enterprise features such as connectors, skills, and experts. The repository is at https://github.com/hikariming/dsh-ee.

It does not run models. Inference and tool calls still happen in each employee's local DSH; dsh-ee manages the capabilities, connections, data, models, and policies that DSH needs. When employees open DSH, they see company-reviewed micro-apps, experts, and connectors; they do not have to search online for plugins or copy and paste policy documents.

Specific features make a long list: the capability hub manages plugins, skills, prompts, enterprise connections, and micro-apps, each with versions, access scopes, and delivery methods; a micro-app consists of a persona, a workbench, a knowledge base, connections, and skills, and employees can open it by typing /name in DSH; enterprise connections go through an MCP gateway, so ERP, CRM, and ticketing systems can be connected, with credentials encrypted and kept server-side and never sent down to employee machines; the data hub queries shared datasets using read-only accounts; the knowledge base does not use a vector database, relying instead on multi-round keyword retrieval and citing the original text; the built-in npm registry includes review that checks install scripts, dependencies, licenses, and leaked secrets; the model gateway is compatible with the OpenAI protocol, can connect to DeepSeek and other upstreams, and supports quotas and failover; and it is a single static Go binary that can run on OceanBase or TiDB and works fully offline.

dsh-ee 一分钟演示:连接企业系统与知识,组装微应用和专家,按部门审核授权,交付到每个员工的 DSH,并审计每次调用

dsh-ee 架构:现有系统与公司资产流入 dsh-ee,组装成微应用,在员工的 DSH 里运行

The demo includes two scenarios: a 200-person sample company, Galaxy Tech, and a four-person fiction studio, Shanhai. You can choose an account manager, expense accountant, new employee, or IT lead and see what each gets in DSH. There is also an admin view showing how systems are connected, how apps are reviewed and authorized, and how each call is logged.

To try it directly, open demo.hikariming.com. If you already have DSH, run a one-line install script, confirm the device code in the browser, then type /customer-brief to ask for a visit brief on a customer.

One commenter said the trio of connectors, skills, and experts is exactly right. What enterprises need is the ability to bring in data, experience, and roles; the model itself is not that scarce. Open-sourcing it lets others build directly on top.

OwnDsh: Manage Identities, Keys, Quotas, and Auditing First

Another project, OwnDsh, takes a different direction. It positions itself as a self-hosted team control plane for DeepSeek Harness. Administrators centrally manage identities, models, API keys, access permissions, quotas, plugins, devices, and auditing; employees continue to work in their own DSH Desktop or Harness Web. The repository is at https://github.com/boe1900/owndsh.

OwnDsh 动态小鲸鱼图标

OwnDsh does not fork the official Harness Web UI, does not take over employee workspaces, and does not remotely execute employee tools. Employees install the standard Harness plugin owndsh-plugin, or use OwnDsh Desktop with the plugin and runtime preinstalled. Employee devices never receive vendor API keys; every model request is revalidated for identity, authorization, and quota by the OwnDsh gateway.

Deployment uses Docker Compose. By default it is accessible at localhost:8080, with initial account admin and password owndsh; the first login forces a change to a real password. Identity supports LOCAL, LDAP, and OIDC. Model administrators centrally store vendor keys, create model sets, authorize them for everyone, user groups, or specific members, and then configure token quotas, RPM, and concurrency. Plugin administrators register npm package names and versions, publish enterprise plugins, and employees discover and install them on the official Plugins Discover page.

OwnDsh is currently in pre-release. Docker images are published only as next, and the npm plugin is also published only as next; latest is not updated. It is not affiliated with DeepSeek AI or Anywhere Labs and integrates only through public plugin extension points.

Two Routes: Choose Based on What You Lack

dsh-ee is more like a capability hub: it assembles company systems, knowledge, experts, and micro-apps and delivers them to employees' local DSH. OwnDsh is more like a control plane: it manages identities, model keys, quotas, plugins, and devices so employees log in with enterprise accounts and never touch upstream credentials.

Both projects are open source or open pre-release, and both sidestep the same problem: enterprises are reluctant to hand internal systems directly to AI. The model itself is not the hard part. Permissions, data, and auditing are. When choosing, first determine whether you lack a way to assemble systems and knowledge into something employees can use, or a way to manage identities, keys, quotas, and auditing.

dsh-ee 微信交流群二维码

发布时间: 2026-11-18 09:37