A plugin-based workbench for AI
DeepSeek Harness (dsh) is an open-source agent framework: models, tools, filesystem, commands, sandbox and approvals are all replaceable plugins, so AI actually gets work done on your machine.
Plain words first
A "harness" channels power where it needs to go. This harness is a runtime for large language models: the model does the "thinking", the harness does the "doing".
Think of it as a workstation for an AI employee: it reads and writes your files, runs commands, browses the web, delegates to subagents, acts only with your approval, and logs every step. You give tasks and keep the gates.
The core idea: everything is a plugin
Traditional software is a welded building; the harness is LEGO. The model adapter, tool registry, session log, agent loop — every block is a plugin, replaceable from configuration.
Traditional monolith
- Features welded into the core
- Swapping models/sandboxes waits on the vendor
- Extending means forking or patching
- Removing a feature means editing source
Plugin-based harness
- Every capability is a removable plugin
- Model, sandbox, approvals are swappable
- Extending = mounting a plugin
- Unloading a plugin unwinds its effects
Powered by Cordis
The plugin composition is driven by the Cordis meta-framework:
Plugins
The smallest unit implementing a capability, mounted on a shared context
Services
Capabilities on ctx, looked up by key
Events
The extension points where plugins talk
Three entry modes
Web GUI
Chat, approve and manage workspaces in the browser — the daily driver
dsh webHeadless
One-shot tasks without a UI, for scripts and CI
dsh --profile headless "task"Plugin management
Install and remove third-party plugins per profile
dsh plugin --profile <name>Got it in a minute — what next?