Unified model access
A consistent interface across DeepSeek Dense, MoE, Code, Math and Vision models.
Developer preview now open
DeepSeek HarnessEverything is a plugin
Models, tools, skills, sessions, sandboxes, storage, loops, orchestration and UI are all composed as plugins—free to replace and recombine.
Open source · Composable · Traceable
DeepSeek Harness Runtime
onlineBuild and verify the current task
agent-loop / run-0042 / workspace
Running
Understand the task
01Read the goal and repository context, then create an executable plan.
Call tools
02Search files and trace the real code path that needs to change.
Verify the result
03Check the change and keep iterating until the delivery target is met.
›harness run --trace --sandbox isolated
Unified execution interface
A backend-agnostic runtime abstraction
DeepSeek Model
Reasoning, code, math and multimodal models
Harness Runtime
Context, loops, plugins and tracing
Real Environment
Repositories, filesystems and enterprise tools
Model × Environment
The model is the soul of an Agent.
Harness gives an Agent the ability to understand its environment, use tools and keep working in real scenarios—so a model can complete tasks, not just answer questions.
Unified model access
Observable execution
Isolated runtime
Configuration driven
Inference deployment
Supports Transformers, vLLM, TGI, llama.cpp / GGUF, TensorRT-LLM and SGLang.
Quantization and compression
Integrated GPTQ, AWQ, FP8, INT8, INT4 and KV Cache quantization strategies.
Evaluation system
Connects general, code, math, reasoning, Chinese, long-context and multimodal benchmarks.
Composable Agent Runtime
Everything is a plugin. Every run is traceable.
DeepSeek Harness is the open-source Agent framework connecting foundation models to real execution environments. Every capability is replaceable and every step is observable.
Agent task loop
Automatically cycles through analysis, tool calls, results, corrections and final delivery.
Code and file operations
Read, search, create and modify files with repository-aware context.
Skills extensions
Package team standards, domain knowledge and repeatable operations in SKILL.md.
Fully pluggable architecture
Models, tools, loops, storage, sandboxes, workflows and UI can all be replaced.
Agent task loop
trace · 0042Analyze
Understand
Call
Select tools
Observe
Read results
Revise
Update plan
Deliver
Verify done
[observe] Tool returned 3 relevant files; call paths mapped.
[reason] Boundary condition found; implementation updated for another pass.
[deliver] Static checks passed; artifact and trace saved.
Design Principles
Design principles
The engineering access layer + capability verification layer + deployment acceleration layer for DeepSeek models.
Backend agnostic
Switch between model services and inference engines through a stable abstraction layer.
Configuration driven
Configure models, quantization, parallelism and plugin policies in YAML / JSON.
Reproducible
Fix random seeds and preserve complete environment information and execution traces.
Extensible
Add models, evaluations, tools and inference engines as plugins.
Use Cases
Built for real work
From a single coding task to enterprise engineering workflows, run continuously on the same composable runtime.
AI coding assistant
Analyze repositories, modify code, run checks and keep fixing based on the results.
use case · 01
Automated code review
Inspect changes, potential defects and conventions, then produce actionable review feedback.
use case · 02
Testing and repair
Read errors, find root causes and iterate on fixes with validation inside the task loop.
use case · 03
Enterprise engineering agents
Connect internal systems through plugins and build assistants aligned with organizational processes.
use case · 04
Multi-agent collaboration
Split large tasks across specialized agents while keeping context and results traceable.
use case · 05
Workflow automation
Package repeatable team processes as Skills or Workflows for reuse.
use case · 06
Plugin Plaza
Compose only what you need
From models to UI, every slot connects through a clear contract. Choose the plugins you need and build your own Harness.
model
Model Provider
Unify DeepSeek API, local inference and private deployments.
tool
Tool Registry
Discover, authorize and call terminals, files, browsers and enterprise tools.
skill
Skills Runtime
Load reusable domain knowledge, team conventions and standard workflows.
runtime
Sandbox
Control commands, filesystems and network boundaries with isolation policies.
memory
Memory Store
Persist task knowledge and recall the right context when it matters.
observe
Trace Viewer
Inspect every inference, tool call, result and state transition.
Standard plugin contracts let team capabilities accumulate and remain replaceable.
Start composingOpen Source · Developer Preview
Build your own DeepSeek Harness
Start with the smallest runtime, then add models, tools, Skills, sandboxes and storage as needed. Bring the Agent into your real work environment.
terminal
Clone the developer preview
$git clone https://github.com/deepseek-ai/deepseek-harness.git
Ready to compose your first Harness