DeepSeek Harness: everything is a plugin

What DeepSeek Harness actually is

The README’s one-liner is everything is a plugin, and DeepSeek’s own product page enumerates what that covers: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all replaceable from configuration. There is no privileged core to patch — the harness runs on Cordis, a plugin kernel where plugins contribute services, typed events, and reversible effects to a shared context, and every part of the product is a plugin mounted beside the others. Want a different model? Register an adapter on the model seam. Different tool surface? Swap the registry. A different orchestration logic entirely? The loop is a plugin too.

The release facts, from the repository itself: CLI name dsh, MIT license, Node.js runtime, launched August 13 2026 — the same day DeepSeek declared V4-Pro generally available. It is a developer preview: the root package reads 0.1.0-rc.5, and the README warns in capital letters that there will be compatibility-breaking changes. Install it with npx @deepseek-ai/dsh web and it serves a browser UI at 127.0.0.1:3080; there is no hosted service — it runs on your machine.

The adoption curve is the fastest the harness category has seen. VentureBeat’s same-day snapshot on August 13 counted roughly 27,500 stars; the count read about 89,000 by August 14; by mid-September 2026 the repository sits north of 212,000. Those are point-in-time snapshots, not a stable metric — but the direction is not in dispute.

The four modes — who each one is for

Profiles stack bundles over those modes: dsh web for the browser UI, dsh --profile headless "job" for a one-shot, listen-on-no-ports run built for CI and scripting, and a community TUI profile installed through dsh plugin — the in-tree TUI package was pulled the week before launch and the ecosystem replaced it in under a week without a fork, which is the plugin architecture working exactly as advertised.

The append-only session log is the real design decision

Every prompt, model response, tool call, tool result, and permission decision is an event in an ordered stream. State is derived, never stored: resume is replaying the log to its tip, fork branches a new session from any event, replay re-runs a trajectory for debugging, and the transcripts, telemetry, and web UI are all projections over the same stream. Engineers will recognize event sourcing ported to agent sessions — and what it buys: time-travel debugging, cheap branching, and an audit trail that is the system of record rather than a lossy export of one.

The die-off-era entrance exam for any agent tool — where do sessions live, in what format, and can you get them out without the vendor’s help? — has its best answer yet here. The record of what your agents actually did, what they touched, and what it cost is the asset that appreciates: your context corpus, your private benchmark, your audit trail.

Migration cost from Claude Code: close to zero

DeepSeek Harness meets the existing ecosystem in three places. It ships optional subagent providers for Claude Code and Codex (disabled by default; each plugin pins and ships its own SDK, and authentication stays native to each product) — the outer loop can run on an open-weight model while the two tasks where a frontier model earns its price get delegated to Claude Code, all recorded in one trajectory. It ships hook bridges that run your existing Claude or OpenAI hooks.json and reads CLAUDE.md / AGENTS.md for context. And it speaks MCP as a client, so your existing servers just work.

Sandboxing is OS-native on all three desktop platforms: Landlock on Linux, Seatbelt on macOS, a restricted-token runner on Windows. Governance is unusual and worth knowing before you invest: the project accepts no external pull requests — community input flows through GitHub Discussions and ships as plugins, which the project says carry equal weight with DeepSeek’s own.

The marketplace question — and the answer that pays

Creator mode plus equal-weight community plugins is a distribution mechanism waiting for a storefront, and DeepSeek ships none: no plugin marketplace, no revenue share, no way for the person who authored the preset that saves your team ten hours a week to earn from it. The architecture proves creators can build on the harness — it does not pay them.

VibeFuse is the live version of that missing layer. It is the first ever free widget-based AI harness: every agent, tool, and CLI runs as a widget on one canvas, processed locally, working in any app. And it runs an open-source marketplace where creators make money selling widgets, skills, and styling packs — 80% revenue share, no listing fee, no new-seller cap. Where DeepSeek Harness asks what a harness could be if every layer were swappable, VibeFuse answers the next question: what happens when the people who build those layers get paid.

Pair them: run dsh headless for CI-grade agent runs, and build the surfaces you actually want to sell — or buy — on VibeFuse. Local voice work while you build? VocalFuse handles dictation and transcription locally for less than a coffee a month.

Go deeper