DeepSeek shipped the harness: dsh makes the runtime, not the model vendor, the choice

DeepSeek published DeepSeek Harness (dsh) on 13 August 2026, the same day it put DeepSeek V4-Pro on its API ^[1][2]. The interesting artefact is not the model: dsh is an MIT-licensed execution runtime that handles the session loop, tool calls, permission prompts and a local web UI, which shifts the thing a team actually selects from the model vendor to the scaffold wrapped around it ^[3][4].

The division of labour is stated plainly in the project's framing: the model predicts tokens, and the harness decides what context and files the model sees, which tools the agent may call, how file writes and shell commands get approved, and how multi-step sessions are stored and replayed ^[5]. Claude Code, Codex CLI and Gemini CLI are harnesses too, each wrapped around one provider's model ^[6], and VentureBeat framed dsh as the open-source competitor to Claude Code ^[7]. According to InfoQ, dsh is built on the Cordis meta-framework with a micro-kernel architecture in which model adapters, tool registries, sandboxing environments, session state handlers, event dispatchers and user interfaces load as isolated, interchangeable plugins ^[8]. Switching model endpoints, remote API or local runtime server, is a declarative YAML or JSON config change rather than a change to core logic ^[9].

That is load-bearing because dsh ships built-in providers for Anthropic, OpenAI, Amazon Bedrock, Google Vertex and Azure, with DeepSeek as default and V4-Pro as the model launched alongside it ^[10][11]. InfoQ reads the release as part of a shift toward modular, unbundled agent infrastructure, separating agent loop, tooling and backend models into decoupled layers ^[12].

Operationally it is small. `npx @deepseek-ai/dsh web` serves the UI at http://127.0.0.1:3080 ^[13]. The API key is set in Settings and stored in `$DSH_HOME/.credentials.yaml`, with the main config holding only a reference to it ^[14]. Choosing a workspace is mandatory, because dsh needs the permitted file scope before a session starts ^[15]. Writes and shell commands raise approval prompts according to the active permission policy ^[16]. `dsh web` is shorthand for `dsh --profile web`, profiles live under `$DSH_HOME/profiles/<name>`, and automation uses a headless profile that runs one session, prints the result and exits ^[17][18]. The 0.1 preview ships four baseline configurations: Standard with shell execution and web retrieval, Code exposing an SDK for batched multi-step tool calls, Minimal restricted to a persistent shell and text editing, and Creator for diagnosing plugin configurations ^[19].

The append-only event log is the part worth attention from anyone who has debugged an agent in anger: every user message, tool invocation, intermediate reasoning state, token metric and sub-agent dispatch lands in one execution trajectory, supporting replay, error isolation and benchmarking model behaviour across runs ^[20].

The price of the plugin surface is more places to break. The README states there will be breaking changes, and InfoQ notes extension contracts and schemas remain subject to them during the preview ^[21][22]. The dev.to walkthrough advises pinning versions and testing in a sandbox, and not binding production workflows to the API or plugins without a version-migration plan ^[23].

Uptake was immediate: the repository showed roughly 169,000 stars and 18,100 forks by 20 August, within a week of release ^[24], an average near 24,000 stars a day ^[25].

Watch three things. InfoQ ties adoption to plugin ecosystem stability, long-term API maintenance and integration with existing developer workflows ^[26]. Developer discussion on Reddit's LocalLLaMA and GitHub Discussions has centred on reactive lifecycle management and dynamic plugin registration, which is where preview breakage will show first ^[27]. And on cost, DeepSeek has made its off-peak discount permanent ^[28].