Skip to main content
Secure Code Sandboxes & Agent RuntimesOpen SourcePlugin-Native Agent Runtime

DeepSeek Harness

Compose secure AI agents from replaceable plugins, tools, and runtimes

An open-source, plugin-native agent runtime from DeepSeek AI with a CLI, Web UI, Python SDK, tool registry, persistent sessions, sandboxing, and approval policies.

Quickstart Execution

Terminal

Get started with DeepSeek Harness in your project.

# Requires Node.js\nnpx @deepseek-ai/dsh web\n\n# Open http://127.0.0.1:3080, select a workspace,\n# then add your DeepSeek API key under Settings → Models.

About DeepSeek Harness

DeepSeek Harness is a composable runtime for building and operating agents that can inspect files, edit workspaces, execute shell commands, call structured tools, delegate work, and maintain plans. Its Cordis architecture implements models, tools, skills, sessions, storage, sandboxes, scheduling, the agent loop, and user interfaces as replaceable plugins coordinated through typed services and events. Sandbox backends, scoped tool restrictions, guarded execution pipelines, and configurable approval policies help constrain filesystem, shell, and network operations.

DK

Danish Product & Data Recipes

How developers and agents use DeepSeek Harness to connect with Danish digital infrastructure.

Build a DAWA address-validation agent

DAWA

Create a scoped tool plugin that queries the public DAWA address API, validates the returned JSON, and exposes normalized Danish addresses to the model.

export DEEPSEEK_API_KEY="sk-..."\nnpx @deepseek-ai/dsh --profile headless "Use curl to query https://api.dataforsyningen.dk/adresser?q=Rådhuspladsen%201 and return canonical address as JSON."

Create a guarded CVR company lookup profile

CVR / Virk Data

Package an organization-approved CVR client as a schema-validated Harness tool, keeping credentials outside model context.

npx @deepseek-ai/dsh web --patch ./profiles/cvr/cordis.patch.yml

Key Capabilities

Plugin-native Cordis architecture for replacing models, tools, sessions, storage, sandboxes, loops, and user interfaces
CLI profiles, local Web UI, headless execution, and an embeddable Python SDK
Schema-validated model-facing tool registry with scoped allowlists, restrictions, cancellation, and execution policies
Persistent agent sessions with append-only logs for tool calls, context injection, replay, recovery, and branching
Filesystem and shell capabilities protected by sandbox backends and configurable approval policies
Composable profiles and bundles for distributing custom agents and third-party integrations without modifying the core