What is dsh?
DSH, short for DeepSeek Harness, is an open source AI agent framework released by DeepSeek AI, currently in developer preview with active iteration. Compatibility-breaking changes are expected. It gives any language model a workspace, a tool registry, a sandbox, and a session log to operate from. Without a harness, a model can only emit text. It cannot open files, run commands, or remember what it did two turns ago.
DeepSeek built DSH for developers building agent harnesses on open-source, reusable, and composable infrastructure. Every agent capability inside it, including the model adapter, tools, skills, sandboxes, sessions, storage, loops, scheduling, and the UI, is a Cordis plugin. Any of them can be swapped, extended, or replaced through configuration without touching core source code. DSH suits platform teams assembling internal automation, model researchers who need a stable benchmark environment, and plugin authors building on top of the open source AI agent framework.
DeepSeek Harness Video
Features & Benefits
- Plugin-first architecture: Swap any agent capability, including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI, by changing configuration rather than core source code.
- Four agent presets: Standard delivers a full coding toolset with file editing, shell, web search, skills, planning, goals, subagents, and workflows. Code exposes tools through the Code Mode SDK so the model can combine multi-step operations in a single TypeScript program. Minimal strips the environment down to persistent bash and str_replace_editor for clean model benchmarking. Creator adds runtime inspection and plugin experiments for developers building custom presets.
- Append-only session log: Records every system prompt, reasoning step, tool call, result, subagent scheduling event, and context injection in DeepSeek Harness so any run can be resumed, forked, searched, or replayed from the same event stream.
- Multi-provider model support: Connect DeepSeek models, catalog providers, or any custom OpenAI-compatible endpoint through the open source AI agent framework’s adapter layer.
- Sandbox policy control: Choose read-only, workspace-write, or full-access execution on Linux, macOS, and Windows backends without coupling permission scope to the rest of the AI agent.
- Cordis kernel: Mounts, unmounts, and reconnects plugins with reversible effects so unloading a plugin cleanly removes all its registrations.
- Capability seams: Define swappable capabilities through Service Definition, Service Provider, and Consumer roles so one configuration change moves an entire capability.
- Patch layer system: Apply configuration overlays by profile, machine, or command-line argument. Later layers win per row without deep-merging keys.
- Community plugin ecosystem: Discover and install third-party plugins via the DSH-plugin topic on GitHub. Categories include tools, browser control, terminal front ends, agent loops, and sandboxing.
- Trajectory view: Inspect every model-visible record by source in a dedicated UI, including system prompts, reasoning, tool results, and context injections.
- One-command install: Launch the local Web UI at http://127.0.0.1:3080 with a single npx command. No invite or waitlist required.
What can DSH do?
- Build a custom AI agent runtime
- Swap AI model providers by configuration
- Run coding tasks in a sandboxed environment
- Compose multi-step tool calls in a single TypeScript program via the Code Mode SDK
- Fork and replay agent sessions
- Benchmark models against a fixed two-tool environment
- Install and author Cordis plugins
- Control file and process permissions per agent run
- Resume interrupted agent sessions
- Inspect full agent trajectory and context
Real-World Applications
A product engineering team can use DSH to build a repository-specific coding agent from a shared preset. Approved models, tools, sandbox policies, and storage choices go into the baseline. Each repository patches only the rows it needs. The open source AI agent framework stays consistent across the org while individual projects keep the flexibility to deviate where it matters.
Clean benchmark signals are hard to get when the harness changes between runs. Researchers comparing different models can lock DSH into Minimal mode, which narrows the runtime to persistent bash and str_replace_editor. The tool surface stays identical across every run, so behavioral differences trace back to the model rather than the environment.
DevOps and platform teams assembling repeatable internal workflows might find the scheduling and approval plugin categories most relevant. Tools, approval steps, storage backends, and run schedules can all be wired into one agent loop. Because every capability is a plugin, the full setup stays inspectable long after it was first assembled.
Plugin authors have a shorter path to publishing with Creator mode available during development. Load a new Cordis plugin in memory, run it against real tasks, and watch what the model sees at each step through the trajectory view. The same session log powering Standard mode is active throughout, so nothing about the testing environment is artificial.
DSH also suits general experimentation for developers who want to try different models on the same task without changing the tool surface. Swap a model adapter in configuration, rerun the session, and compare trajectories side by side using the fork and replay features.
