OHarness vs DeepSeek Harness

Choose DeepSeek Harness if you want to take the agent runtime itself apart — it is MIT licensed, the source is public, and every piece is a plugin — and can live with a developer preview. Choose OHarness if you want a coding agent to use every day, on any model, in the terminal, browser or desktop.

DeepSeek Harness’s side checked against its own documentation on 2026-09-25. Sources are listed below.

Side by side

	OHarness	DeepSeek Harness
Form factor	Terminal UI, browser UI, desktop app (macOS) and a headless runner — one engine	Local web UI (`dsh web`), an Electron desktop app built from source, a CLI with a headless profile, an ACP server, and TypeScript and Python SDKs
Licence and source	Free to use; npm packages MIT. Source repository not public yet	MIT, source public
Models	Anthropic, OpenAI and Gemini natively; any OpenAI-compatible endpoint; 30+ catalogued providers	DeepSeek directly, built-in providers such as Anthropic, OpenAI, Kimi and GLM, and custom endpoints speaking Chat Completions, Responses or Anthropic Messages
Local models	Ollama, LM Studio and vLLM built in, no key	Through a custom OpenAI-compatible endpoint to a self-hosted server; vLLM settings are documented
Switching mid-session	`/model` mid-session; history is vendor-neutral, nothing converted	`/model` or the model picker; the change applies from the next request
MCP	Yes — stdio and Streamable HTTP, with OAuth for remote servers	Yes — stdio and Streamable HTTP, opt-in per server
Permissions	Five modes plus allow/ask/deny rules in user, project and local scopes; deny beats bypass	Presets pairing a sandbox mode with an approval policy (`ask` or `never`); the default is workspace-write with ask
Sandbox	No OS sandbox — a policy gate on tool arguments. Use a container for confinement	OS-enforced for file writes: bwrap and Landlock on Linux, Seatbelt on macOS, restricted tokens on Windows; network is outside its scope
Subagents	Yes — isolated contexts, run in parallel	Yes — in-process, or delegated to ACP, Codex, Claude Code or SDK agents; agent teams are experimental
Headless	`oharness-headless`, JSONL on stdout	`dsh --profile headless` prints the answer; `--json` streams newline-delimited JSON events
Maturity	Young — 0.8 on npm and a 0.3 macOS desktop build — but released rather than labelled a preview	Developer preview since August 2026; its README warns of compatibility-breaking changes and its safety notice says it has had no security audit
Price	Free with your own key or a local model; optional hosted plan	Free under MIT; model usage is billed by the provider

Where DeepSeek Harness is stronger

The source is public under MIT today; OHarness's is not yet.
Everything is a plugin — the model adapter, tools, agent loop, sessions, sandbox and UI can each be swapped.
An OS-level sandbox for file writes on Linux, macOS and Windows. OHarness has a policy gate, not confinement.
It can hand subtasks to other agents over ACP, Claude Code and Codex among them.
It is DeepSeek's own project, so DeepSeek models and account sign-in are first-class.

Where OHarness is stronger

Built to be used every day rather than restructured: a terminal UI, a browser UI and a desktop download, all on one engine and one set of sessions.
Ollama, LM Studio and vLLM are built-in providers that need no configuration and discover their models; DeepSeek Harness reaches local models through a custom endpoint.
Conversation history is stored vendor-neutral, so a mid-session switch between vendors converts nothing.
deny rules hold even in bypass mode, and rules stack across user, project and local scopes.

Which to choose

DeepSeek Harness, if you want an MIT codebase you can read and rebuild today, an OS sandbox for writes, or an agent that orchestrates other agents over ACP — and preview-stage breaking changes do not worry you.

OHarness, if you want a coding agent for daily work in the terminal, browser or desktop, with local models built in and cross-vendor switching that converts nothing.

Common questions

Does DeepSeek Harness only work with DeepSeek models?

No. It has a direct DeepSeek adapter, built-in providers such as Anthropic, OpenAI, Kimi and GLM, and custom endpoints in three API formats: Chat Completions, Responses and Anthropic Messages.

Is DeepSeek Harness ready for daily use?

By its own description it is a developer preview: the README warns of compatibility-breaking changes, the safety notice says it has not had a security audit, and its npm releases are 0.1.x release candidates as of September 2026.

Can I use DeepSeek models in OHarness?

Yes. The deepseek provider speaks DeepSeek's API with a DEEPSEEK_API_KEY, and V4 Flash and V4 Pro are in the catalogue. The DeepSeek V4 guide covers setup and what a session costs.

Sources

Spotted something out of date? Tell us and it will be corrected. More comparisons are on the guides page, and Claude Code alternatives covers what to check before switching.