- CollectionAgent Harnesses & Operator Surfaces
- Classificationagent-native
- LicenseMIT
- MCPUnavailable
- Agent SkillUnavailable
- InterfacesCLI · SDK · ACP · json rpc · http
- Verified2026-10-13
What it does
DeepSeek Harness (dsh) is DeepSeek AI’s open-source agent harness. Official positioning is that a model is the agent’s soul and the harness is what lets the agent understand its environment, use tools, and keep working in real scenes. Every agent capability — model, tools, skills, session, sandbox, storage, loop, scheduling, UI — is a Cordis plugin. Developers compose or replace those plugins in configuration without forking the harness.
The product is a developer preview. The README warns there will be compatibility-breaking changes. The homepage lists four run modes: Standard (full coding agent), PTC (standard tools plus a Code Mode TypeScript program that batches tool calls), Minimal (persistent bash + str_replace_editor for model benchmarks), and Create (inspect the runtime, experiment with plugins, author new presets).
Agent-facing interfaces
- Interaction pattern: CLI + Web UI (developer preview)
- Catalog protocols: CLI, SDK, ACP, json rpc, http
- MCP: ⚠️ Not published as a standalone MCP server.
- Agent Skill: ⚠️ No official npx skills add package published yet.
- Onboarding type:
cli
How agents onboard
Interaction pattern: CLI + Web UI (developer preview)
npx @deepseek-ai/dsh web
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| CLI / npm | npx @deepseek-ai/dsh web; source pnpm dsh |
| Web UI | Local operator UI, default http://127.0.0.1:3080 |
| ACP | Automation-only Agent Client Protocol server |
| JSON-RPC SDK | TypeScript client in packages/sdk/ |
| Python SDK | Bundled runtime under python/ |
| Session log | Append-only events; Trajectory resume/fork/replay |
| Plugin topic | dsh-plugin for community plugins |
Verification
Last verified 2026-10-13.
Sources:
- https://api.github.com/repos/deepseek-ai/deepseek-harness