|

DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin

ByRicardo

DeepSeek launched DeepSeek Harness v0.1 in developer preview and printed the complete supply code beneath the MIT license. The challenge ships as dsh at deepseek-ai/deepseek-harness. A harness is the layer between a mannequin and the atmosphere it acts in — the instruments, recordsdata, sandboxes, and management loop that allow an agent hold working. DeepSeek frames it as Agent = Model + Harness. Most harnesses hard-code that layer: the agent loop, the device registry, and the session retailer are mounted, and extension occurs solely at no matter hooks the authors uncovered. Harness takes the other place, acknowledged in the primary strains of its README: the whole lot is a plugin. Models, instruments, expertise, classes, sandboxes, storage, loops, scheduling, and the UI all sit behind Cordis plugin boundaries, and any of them might be chosen, swapped, or prolonged in configuration with out altering Harness supply code. That makes this a equipment for assembling agent runtimes somewhat than a mounted coding assistant, and it is why the discharge issues greater than the mannequin announcement it shipped alongside.

Is it deployable?

Yes, however as developer infrastructure, not as a manufacturing agent product. v0.1 is a developer preview.

  • Company stage: AI-native startups and platform or developer-experience groups inside mid-to-large enterprises that already run inside tooling. Regulated enterprises can pilot it domestically as a result of it is MIT-licensed and self-hosted.
  • Industries: Software and devtools, monetary providers and insurance coverage (auditable agent runs), healthcare and pharma R&D, cloud and semiconductor distributors publishing reference agent stacks, and tutorial or industrial analysis labs benchmarking fashions.
  • Applications: Internal coding brokers over personal repositories, mannequin analysis inside a managed two-tool atmosphere, agent observability and run replay, customized sandbox and approval insurance policies, and packaging home tooling as reusable plugins.

The Cordis kernel

Harness runs on Cordis, a meta-framework whose design is set out in A Programming Paradigm for Spatiotemporal Composability. The kernel handles plugin mounting, unmounting, and dependencies. Capabilities reside in the plugins, not in a privileged core.

Plugins cowl fashions, instruments, expertise, classes, sandboxes, storage, loops, scheduling, and the UI. Cordis providers and occasions allow them to work collectively. Developers choose, swap, or prolong any functionality in configuration, with out altering Harness supply code.

Four runtime modes

Standard is the complete coding agent: file enhancing, shell, file and net search, expertise, planning, targets, subagents, and workflows. Code mode exposes these instruments via a Code Mode SDK, so the mannequin can mix multi-step operations in one TypeScript program. Minimal retains two instruments, a persistent bash and str_replace_editor, for benchmarking fashions in a naked atmosphere. Creator mode provides runtime inspection, in-memory plugin experiments, and preset-authoring steering.

Every run is traceable

Everything the mannequin sees is written to an append-only session log. That consists of system prompts, reasoning, device calls and outcomes, subagent scheduling, and each context injection. The Trajectory view inspects these data by supply. Resume, fork, search, and replay all function on the identical occasion stream. Most agent frameworks log device calls; recording each context injection is the sharper declare right here.

Model routing is additionally a plugin

In Settings → Models, a DeepSeek API key takes impact on the following request with out restarting the server. The installed catalog provides suppliers reminiscent of Anthropic and OpenAI by API key. Bedrock, Vertex, Azure, and Codex want native credentials as an alternative: AWS credentials and a area, an ADC challenge, an api-version, and OAuth respectively. Custom suppliers settle for any OpenAI-compatible base URL and protocol. Keys are write-only and saved in $DSH_HOME/.credentials.yaml; settings retain solely a credential reference.

Running it

npx @deepseek-ai/dsh net begins the Web UI, served at http://127.0.0.1:3080 by default. From a checkout: git clone, pnpm set up, pnpm run construct, then pnpm dsh net. A Python SDK ships as deepseek-harness-sdk and wishes Python 3.10 or newer, on Linux x64, Linux arm64, or macOS 14+ on arm64. Its bundled runtime requires no system Node.js.