BISHENG: a 12k-star open enterprise agent platform with a DeepSeek Harness desktop client

Enterprises adopting agents need more than a chat window: orchestration, knowledge bases, permissions, auditing and private deployment as one package. BISHENG, open-sourced by Chinese team dataelement (Apache-2.0, 12,145 stars as of November 6, 2026), is built to exactly that complete checklist — agent and workflow orchestration, RAG and knowledge management, unified model management, evaluation, observability, access control and enterprise governance. The official positioning: an open-source alternative to WorkBuddy / Doubao Work, keeping AI capabilities in the enterprise’s own hands rather than a vendor’s cloud.

Core features

• Desktop client on DeepSeek Harness (DSH): works with the BISHENG platform as an open, privately deployable suite; locally developed apps go through pre-checks, secret scanning, build and release approval before landing in the app center.
• Task-mode general agent: the main agent generates and tracks a task list dynamically, delegating independent research to parallel sub-agents.
• Enterprise governance: unified model management, evaluation, observability, access control and an admin console with credit and usage tracking built in.
• Platform reuse: apps plug into platform login, visitor identity, models and knowledge resources instead of rebuilding each.
• RAG and knowledge management: built-in knowledge bases wired to enterprise corpora for retrieval and Q&A.

Typical use cases

• Enterprise AI rollout: from PoC to multi-department adoption with permissions and auditing covered by the platform.
• Office automation: agent orchestration for document processing, approval flows and knowledge Q&A.
• Private deployment: the open-source route when data cannot leave the intranet.

Quick start

Docker Compose brings up the platform; the desktop client downloads from dshdesktop.com. Installation and configuration details live in the official wiki (Feishu docs). The division of labor is deliberate: heavy orchestration and governance stay on the server while task-running agents sit on each employee’s machine, tied together by the platform’s account system. Six-language READMEs (Chinese, English, Japanese, Spanish, Portuguese, Russian) and a Trendshift badge hint at the community scale, and upstream maintains dedicated evaluation and documentation sites, so onboarding material is complete.

Summary

BISHENG fits mid-to-large organizations building or privately deploying an enterprise agent platform; individual lightweight use and pure chat scenarios don’t need it. Facts and limits: Apache-2.0, Python-led stack, with industry leaders and Fortune 500 organizations among claimed users (official figure); some enterprise governance capabilities sit in the commercial tier — check version notes before self-hosting. It competes with Dify and Coze-class platforms, differentiated by the desktop-plus-DeepSeek-Harness local route and office-scenario depth — where auto-ontology on this site supplies the semantic layer, BISHENG supplies the platform layer.