DeepSeek AI Releases DeepSeek Harness in Developer Preview

DeepSeek AI has announced the release of DeepSeek Harness v0.1 in developer preview, positioned as developer infrastructure rather than a production-ready agent product. The software is distributed under an MIT license and is engineered to provide maximum flexibility for artificial intelligence developers.

The underlying architecture of the system runs on Cordis, a meta-framework designed around the principles outlined in A Programming Paradigm for Spatiotemporal Composability. The kernel manages plugin mounting, unmounting, and dependencies, while specific capabilities reside within the plugins rather than a privileged core.

Stock photo for illustration only, not from the actual event

Plugins cover a wide range of components including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and user interfaces. Cordis services and events enable these components to work together seamlessly, allowing developers to select, swap, or extend any capability via configuration files without altering the core Harness source code.

The core operation modes provided by the system include:

• Standard: A complete coding agent supporting file editing, shell commands, file and web searches, skills, planning, goals, subagents, and workflows.
• Code mode: Exposes developer tools through a Code Mode SDK, allowing models to combine multi-step operations within a single TypeScript program.
• Minimal: Retains only two essential tools—a persistent bash shell and str_replace_editor—for benchmarking models in a bare environment.
• Creator mode: Introduces runtime inspection capabilities, in-memory plugin experiments, and preset-authoring guidance.

Every piece of data observed by the model is recorded in an append-only session log, encompassing system prompts, reasoning steps, tool calls and outputs, subagent scheduling, and context injections. The Trajectory view inspects these historical records by source, while resume, fork, search, and replay functions all operate on the exact same event stream.

Regarding credential management, entering a DeepSeek API key under Settings → Models applies it immediately to the next request without requiring a server restart. The installed provider catalog also accommodates alternatives such as Anthropic and OpenAI via API keys, alongside Bedrock, Vertex, Azure, and Codex which require native credentials. All keys are stored as write-only in $DSH_HOME/.credentials.yaml.

Getting started is straightforward via npx @deepseek-ai/dsh web to launch the Web UI at port 127.0.0.1:3080, or by cloning and building from the source repository. Additionally, a Python SDK named deepseek-harness-sdk requires Python 3.10 or newer on Linux x64, Linux arm64, or macOS 14+ on arm64, featuring a bundled runtime that operates independently of system Node.js.

Source: MarkTechPost