Hermes Agent vs DeepSeek Harness: which agent to use

Hermes Agent and DeepSeek Harness at a glance

Models, tools and plugins

Hermes Agent is not tied to one model. Its provider list includes DeepSeek: set DEEPSEEK_API_KEY and choose the deepseek provider. It supports many other providers and compatible endpoints too, so you can switch models without changing agent. A separate page will cover running Hermes on DeepSeek models in more depth.

DeepSeek Harness is set up around DeepSeek first: in Settings you paste a DeepSeek API key and the route works at once. Its guide also lets you add third-party providers it ships with (for example Anthropic, OpenAI, Kimi and GLM) and custom endpoints that speak the OpenAI Chat Completions, OpenAI Responses or Anthropic Messages protocols. So both can use DeepSeek models, and both can use others.

Both read and edit files, run commands and delegate work to sub-agents. Hermes ships a large set of built-in tools (terminal, files, browser, web search, image generation, voice and more), can use MCP servers, and has a plugin system for custom tools and hooks. It can run commands in Docker or over SSH as well as on the machine itself.

Plugins are DeepSeek Harness’s centre. Its README calls the design “everything is a plugin”, and DeepSeek’s page lists official plugins for agent teams, auto-approval review, scheduled tasks, voice input and a terminal. DeepSeek’s page says you can also create plugins through chat in Creator mode, and an experimental plugin market for installing community bundles. If you want to shape the tool itself, this is where it is strongest. Several of those pieces are marked experimental in the repository.

Memory and skills

Hermes keeps two small files of notes between sessions: what it has learned about its environment and what it knows about you, with a fixed size limit so they stay focused. The agent edits them itself with its memory tool, and past conversations stay searchable. For skills, it can write and improve its own, as covered in how Hermes Agent skills work.

DeepSeek Harness has a skill system too: reusable instructions that are loaded only when needed, found in project or user folders, with Word, PowerPoint and Excel skills available, and invoked by name. In the documentation I read, long-term memory is not a built-in feature; the guides show how to connect a third-party memory server through MCP, using three example setups that are switched off by default and described as interoperability examples, not endorsements.

Chat apps, scheduling and always-on use

This is the clearest difference. Hermes’s messaging gateway is a background process that connects Telegram, Discord, Slack, WhatsApp, Signal, email, WeChat and other apps, and it also runs your cron jobs. A cron job can be one-off or recurring and can use skills. Jobs can deliver their results to a configured chat while the gateway runs; jobs created from a messaging app default to that chat, and jobs created from the command line default to local files. All of this only happens while the computer or server running Hermes is switched on.

DeepSeek Harness is used mainly as a workspace in the browser or the desktop app, and it also does scheduled work. It can create one-off, repeating, daily, weekly and cron-style tasks. While its Host is running, a due task restores the original conversation and queues a follow-up message into it, which the agent then handles. Delivery is into that conversation, not email, text or push notification, and closing the Host stops scheduling until it starts again. I found no chat-app integration in its documentation, and I have not compared how scheduling behaves between its Web and Desktop versions. The repository does include an experimental signed GitHub webhook adapter for deployments that route events in. The Web interface runs at an address on your own machine by default, and its docs cover putting it behind a reverse proxy if you want to reach it elsewhere.

Safety and undoing mistakes

Hermes can ask for approval before dangerous shell commands, depending on approvals.mode. The default, smart, has an auxiliary model assess a flagged command: it can approve a low-risk one, deny a clearly dangerous one or pass an uncertain one to you, and an interactive owner can approve a denied command once. manual always asks, and off disables the checks.

Hermes can also snapshot your project before file changes, but this is off by default. In a session, /rollback lists snapshots and /rollback <N> restores one, normally keeping later hand-edits; --all asks for a full restore. By default files over a configurable 10 MB size limit are skipped, and container backends such as Docker get no snapshots.

DeepSeek Harness applies a filesystem policy to commands: read-only, write inside the workspace (and some platform temporary paths), or unrestricted. It asks before operations that need approval, and a denied action can be retried with a one-time approval from you. An experimental auto-review plugin lets the model assess supported native and inner tool calls first (direct Node effects inside programs fall outside that review), and its README warns it can allow unsafe actions. The project’s safety notice is blunt: it is experimental, has not had a security audit, and should be run with least privilege, ideally in a disposable virtual machine or container, with backups of anything it can reach. Its session checkpoints protect against losing work in a crash; I did not find a file-rollback command in the documentation.

Setup

Hermes installs with one command on Linux, macOS (Apple silicon), WSL2, native Windows and Android (Termux), or through the Hermes Desktop installer on macOS and Windows; see how to install Hermes Agent. Then you configure a model provider and, for chat, set up a messaging app.

DeepSeek Harness needs Node.js (22.19 or newer within the 22.x line, or 24 and above, per its package file) and one command, npx @deepseek-ai/dsh web, which opens the interface at 127.0.0.1:3080. DeepSeek also offers desktop downloads for macOS (Apple silicon) and Windows (64-bit). You then paste an API key and choose a workspace folder. It is quick to try.

What each is for, and which to pick

Hermes Agent is built to be a long-running personal agent: memory between sessions, skills it writes from its own work, scheduled jobs and a gateway to your chat apps. DeepSeek Harness describes itself as an open-source harness built on an “everything is a plugin” design, and DeepSeek’s page shows it handling documents, data analysis, coding help, research and background automation in a workspace you open yourself.

• Pick Hermes Agent if you want one agent that keeps notes on you between sessions, learns skills, runs jobs on a schedule and answers in Telegram, WhatsApp or another app, using a supported model and provider, or a compatible endpoint with working tool calling.
• Pick DeepSeek Harness if you want a plugin-centred workspace built around DeepSeek models on your own computer, and you are comfortable with a developer preview that may change under you.
• Use both if you are curious: they are separate programs, and Hermes can use DeepSeek models too.

Running Hermes on ReadyHermes

If you want the always-on Hermes side without looking after a computer for it, ReadyHermes runs your own Hermes Agent on our servers. It keeps running scheduled jobs and answering your chat apps while your own computer is off, and you can use it from a browser workspace or from Telegram, WhatsApp, Discord, WeChat and others. A new agent is ready in about 10 seconds, and there is a 1-hour free trial.

Paid plans include US$1 of AI a month, or you can connect your own AI provider key. ReadyHermes does not run models on your computer and has no GPU hosting for your own local models. For the trade-offs of running it yourself, see self-hosting vs managed hosting.