Running DeepSeek Harness for the first time

Tested macOS (Apple Silicon) · v0.1 developer preview · verified 2026-08-14

Field note · everything below happened on our machine. Screenshots are from the actual session.

DeepSeek open-sourced the Harness on 2026-08-13 alongside the V4 Pro release, MIT license, v0.1 developer preview. By the evening of 2026-08-14 it had passed 90,000 stars. We walked the entire onboarding on a clean macOS machine the day after release — here is exactly what happened.

Prerequisites

Node.js installed (we ran v22.22.3). Nothing else — no sign-up, no invite code, no payment.
A DeepSeek API key only if you actually want to chat with a model. The UI lets you skip it.

Step 1 — one command

npx @deepseek-ai/dsh web

First execution downloads dependencies; give it a few minutes. When it finishes, the terminal prints dsh web: http://127.0.0.1:3080. Everything runs locally — no account system, no cloud dependency.

Step 2 — the beta consent page

Opening http://127.0.0.1:3080 first shows an internal-preview disclaimer aimed at harness developers. Click Continue.

Step 3 — API key (or skip it)

The setup wizard asks for a DeepSeek API key. Configure later skips straight into the main UI — useful if you just want to inspect the interface.

DeepSeek Harness first-run API key configuration screen — First-run API key configuration. "Configure later" is a fully supported path.

DeepSeek Harness main UI showing New Session, workspace selector and Standard mode — The main screen after skipping the key: a Preview badge, the "Into the Unknown" welcome area, workspace and mode selectors, and the session list on the left.

What the mode selector actually offers

Four modes ship in v0.1, each a different default plugin set:

Standard — the full toolset, default for daily agent use.
PTC — programmatic tool calling: the model writes code that orchestrates multi-round tool calls, cutting round-trips on long chains.
Minimal — one shell tool and one file editor. This is what the official benchmark uses, for clean model comparisons.
Create — inspect the running Cordis runtime and compose plugins in memory.

Things worth knowing before you invest time

v0.1 means breaking changes. The README bolds this. Build against it expecting refactors.
Everything is a plugin. The core is just a Cordis plugin loader; models, tools, skills, sessions, sandboxes, storage, UI are all plugins. Swapping the model plugin and swapping a file editor are architecturally identical operations.
Append-only session logs. Every system prompt, thought, tool call and result goes into a single append-only event stream. Resume, fork, search and replay all share it — this is the most production-minded part of the design.

Two-minute quick start

# start (needs Node.js)
npx @deepseek-ai/dsh web

# open http://127.0.0.1:3080
#   beta disclaimer → Continue → API key (or Configure later) → main UI

# read the source instead
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness && pnpm install && pnpm run build && pnpm dsh web

Verdict: the shortest credible onboarding we've seen in this category — one command, zero accounts, and the UI is honest about being a developer preview. The plugin architecture and the event-stream sessions are worth reading the source for, even if you don't adopt it. For production use, wait a few versions.

Facts verified against: GitHub API (star count, license, release date) on 2026-08-14; official README and docs; and the onboarding session above. Base repo: deepseek-ai/deepseek-harness.