Substructure vs. DeepSeek Harness

DeepSeek Harness (dsh) is for one developer on one machine: an MIT-licensed agent runtime in developer preview since August 2026, built on the Cordis framework, where everything is a plugin: the model adapter, the tools, the skills, the sessions, the sandbox, the storage, the loop, and the UI. Substructure is a server that runs agents for a team: config in git, Slack, MCP authorization, history, and durable runs.

Harness is a client with a shell and a file system. Substructure is a service that clients call.

Use DeepSeek Harness when	Use Substructure when
The agent edits files on your computer	The agent answers your team
You want to replace the loop, storage, or UI	You want the loop open but not owned
One developer runs it	It must run after you close your laptop

What Substructure does that DeepSeek Harness does not

Runs on a server for a team. dsh web serves http://127.0.0.1:3080, and the CLI refuses --host 0.0.0.0, because the agent runs commands and an open port would give that power to the network.
Authenticates people. The Harness web UI has no login. Remote access means a proxy, TLS, and authentication you add yourself.
Restarts a dead run. The Harness log supports resume, fork, and replay, and you decide when to use them. Substructure's engine continues from the last saved step by itself.
Answers in Slack. Harness surfaces are a local web UI, a terminal UI, and a headless mode for scripts and CI.
Takes your code in any language. Harness plugins are Cordis plugins in TypeScript, on Node.js.

Side by side

Building the agent

	DeepSeek Harness	Substructure
The agent is	A profile of mounted plugins	A file in your repo
You extend it with	Cordis plugins and config patches	Config, plus an HTTP worker
Language	TypeScript, on Node.js	Any language, over HTTP
Run modes	Standard, code, minimal, creator	One loop, overridable per step
Skills and plugins	Its own plugin tree	agent-plugins directories
Models	DeepSeek V4, plus a provider catalog	Any provider, one line

Running it

	DeepSeek Harness	Substructure
Where it runs	Your machine, on `127.0.0.1`	A server, cloud or self-hosted
Users	One	Everyone in the deployment
Session log	Append-only; resume, fork, replay	Append-only; the engine resumes
After a crash	You resume the session	The run continues on its own
Status	Developer preview, breaking changes	Stable

Tools and surfaces

	DeepSeek Harness	Substructure
Shell and files	Sandboxed on your machine	Not offered
Sandbox	read-only, workspace-write, full access	Tools run in your worker or an MCP server
MCP	Off until you enable each server	The engine holds the credentials
Sub-agents	Plugins, plus Claude Code and Codex bridges	Other agents in the file
Surfaces	Local web UI, TUI, headless	Slack, AG-UI, CLI, your API

What you write

terminal

npx @deepseek-ai/dsh web        # serves http://127.0.0.1:3080

substructure.toml

[agent.oncall]
llm = "openrouter"
model = "moonshotai/kimi-k3"
system = "You are the on-call assistant."
mcp = [{ id = "sentry", tools = { read_only = true } }]

[slack]
mentions = "oncall"

subs apply, and the bot answers in Slack threads. Add a worker URL and the engine sends every decision to your endpoint as JSON, one step at a time.

Where DeepSeek Harness is the better choice

Editing files on your computer, inside a sandbox. Substructure does not do that.
Replacing the agent loop, the storage, or the interface. The plugin tree goes further than most products allow.
Benchmarking models, with the minimal mode and replay tests.
One local process, with your keys on your own disk.

Questions

Can it run on a server? Only behind a proxy, TLS, and authentication you add. The CLI blocks binding to the network on purpose.

Do both keep a full log? Yes. Harness checks each model call against its log; Substructure uses the log to continue a run after a crash.

Can I use both? Yes. Harness edits code on your laptop; Substructure runs the agent your team talks to.

The quick start takes about five minutes.