DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI, currently a developer preview with compatibility-breaking changes expected. It runs as a desktop app or as a web UI started from code, and it covers everyday office work (organizing material, analyzing spreadsheets, writing documents and reports), coding work (understanding a project, fixing issues, building features, running tests), and research that searches, cross-checks, and cites its sources.
The harness is built on an everything-is-a-plugin architecture powered by Cordis, a framework for composable plugins. Agent capabilities come from plugins that you install from the community or write yourself, and the agent can create and install a plugin from within a conversation. Community plugin repositories publish under the dsh-plugin topic.
Long-running work such as scripts and batch file processing runs in the background with visible progress, and scheduled tasks repeat a prompt on a timetable, for example a weekly project report. A run view breaks each conversation into rounds and steps, with every tool call's parameters, result, schema, and timing.
Features
- Plugin architecture: every capability is a plugin, built on the Cordis framework
- Custom plugins: install community plugins or write your own, including from inside a conversation
- Desktop and web: runs as a desktop application or a browser-based web UI
- Office work: organize material, analyze spreadsheets, and draft documents and reports
- Coding: understand a repository, fix issues, develop features, and run tests
- Sourced research: search, cross-check information, and cite sources
- Background tasks: run scripts and batch jobs in the background with progress tracking
- Scheduled tasks: repeat a prompt on a daily or weekly schedule
- Run inspection: per-round and per-step view of tool calls with parameters, results, and timing
