How does the 7-day trial work?

You get 7 days free of the full PRO plan, with all advanced features: multiple chatbots, multimedia, lead capture, WhatsApp integration and more. We ask for a credit card to start the trial, but we don't charge you during those 7 days. We notify you before charging and you can cancel with one click from your panel anytime.

Do I have to pay anything if I cancel during the trial?

No, nothing. You cancel with one click from your panel and your PRO plan access ends right then. No calls, no forms.

Is there any kind of commitment?

No commitment. You cancel whenever you want and keep access until the end of the period you've already paid for. If you're on the trial, canceling means your access ends right then — but we don't charge you anything either.

Can I upgrade or downgrade my plan anytime?

Yes, from your panel. If you upgrade mid-month, you only pay the prorated difference. If you downgrade, the change applies from the next cycle and you don't lose what you've already paid.

Do I need to know how to code to set it up?

No. You upload documents or URLs and the system learns automatically. Integrating it into your website is one line of code (copy and paste). Connecting WhatsApp is a couple of clicks.

How long does it take to be ready?

Pretty fast. In 5 minutes you have a bot uploading and responding with your basic content. In about half an hour you test it with real questions, see where it falls short, and adjust the content so it answers the way you want.

Can I manage the chatbot from Claude, ChatGPT or similar tools?

Yes. Bravos has its own MCP connector: you connect your account to Claude, ChatGPT or any other AI tool that supports MCP and, from the chat itself, the AI analyzes conversations, audits and improves the content and rewrites the bot instructions, without opening the panel. It is included from the Starter plan and during the trial, and the AI part runs on your own subscription (Claude, ChatGPT or whichever you use), so it adds no cost on Bravos.

What languages does the bot work in?

Spanish, English, French, German, Italian, Portuguese, Dutch, Russian, Polish, Arabic, Chinese, Japanese, and Korean. The bot automatically detects each customer's language and replies in theirs, without you having to configure anything.

What happens if the bot doesn't know how to answer something?

It tells you. If it doesn't find the info in your content, it doesn't make it up: it tells the customer it doesn't have that information and logs the question in your analytics panel. So you can see what's missing and add it in a couple of clicks.

Are my data and my customers' data safe?

Yes. Everything is encrypted in transit and at rest. We comply with GDPR. By default we use OpenAI to generate responses; per their enterprise API policy, they don't store or train models with your data. Enterprise customers can choose other providers (Anthropic Claude, etc.). In any case, your documents and conversations remain private and are not shared with third parties for commercial purposes.

Do I need a new phone number to connect WhatsApp?

No. You can keep the same number you already use in WhatsApp Business. The connection is a couple of clicks from the panel via Meta. Important note: once connected to the bot, conversations are handled via API and you won't be able to respond manually from the WhatsApp Business app with that same number. If you want to keep responding manually in parallel, use a different number for the bot.

What Is DeepSeek Harness? The Open-Source Coding Agent Where Everything Is a Plugin

DeepSeek Harness is DeepSeek’s open-source coding agent: the software layer that wraps an AI model and turns it into an assistant that can read your files, run commands and keep working through long tasks on its own. It launched on August 13, 2026, is free (MIT license), and passed 177,000 GitHub stars in a single week. What sets it apart isn’t the model — it’s the architecture: literally everything (the model, the tools, the interface, even the agent loop) is a plugin you can swap.

This guide explains what DeepSeek Harness is with real detail, not slogans: what “everything is a plugin” actually means (with concrete examples of what the community has already built), how plugins install, its modes, how it compares to Claude Code and Codex, the security risk almost nobody mentions, and how to get started. Sources at the bottom.

What is a coding-agent “harness”

An AI model like DeepSeek, GPT or Claude, on its own, only writes text. For it to actually code — open your files, run the terminal, search the project and keep going until the job is done — it needs a layer of software around it that gives it those hands. That layer is the harness. Claude Code, Codex and Cursor are harnesses. The model is the brain; the harness is the body.

A coding agent is, at heart, a type of AI agent specialized in programming: it doesn’t just answer, it plans steps, uses tools and acts until the goal is met. And here’s why the harness matters so much: the same model performs differently depending on the harness it runs through. The harness decides what the model sees, which tools it has and how much each task costs. That’s why it has become the battleground where AI companies now compete.

What is DeepSeek Harness

DeepSeek Harness (short: dsh) is an open-source coding agent, MIT-licensed, that runs both as a web app and headless (no interface, for automation). Instead of shipping a closed tool with a fixed set of features — like Claude Code or Codex — DeepSeek built the whole harness out of interchangeable pieces, on top of a framework called Cordis designed for exactly that: mounting and unmounting components on the fly.

Note:A bit of jargon: a plugin is a module you plug in and remove to add or change a capability. A language model (LLM) is the AI behind tools like ChatGPT, Claude or DeepSeek. Headless means “no graphical interface”: you drive it with commands or scripts, handy for automation.

“Everything is a plugin”: what it really means

This is the heart of DeepSeek Harness, and it’s worth explaining properly because it’s the whole point. Normally a tool like Claude Code is monolithic: it comes with its model, its interface and its way of doing things, and if you don’t like something, tough. DeepSeek Harness is the opposite: every part of the program is a piece you can remove, swap or replace from configuration, without touching the core code.

“Every part” is literal. The things that are welded shut in other tools are plugins here:

The model: it uses DeepSeek by default, but you can wire in another provider by swapping a plugin.
Tools and skills: browser control, voice, vision, code review, Git integration… all added as plugins.
Memory and sessions: how it remembers and where it stores history is swappable too.
The interface and the look: from the visual theme to the web UI itself.
The agent loop: even the logic of how the agent thinks-acts-repeats is a plugin.

It’s the same idea behind standards like MCP — plugging new capabilities into an agent — only DeepSeek Harness takes it to the extreme: even its own skeleton is replaceable.

“The model is the soul; the harness is what turns it into an agent acting in a real environment: it manages environment interaction, tool use and long-running execution.”
— DeepSeek — Harness (official docs)

Real examples: what the community already built

The best way to grasp “everything is a plugin” is to see what people have already made. In its first week alive, the community shipped hundreds of plugins. Technically, a plugin is a package that declares a dsh.bundle manifest and installs with one command:

dsh plugin add <plugin-name>

A sample of the serious stuff you can already bolt on, to show the range:

Switch models or watch your API spend: dsh-model-picker, dsh-balance.
A real terminal inside the UI (with xterm.js): terminal.
A VS Code-style file explorer with a Monaco editor: dsh-file.
Manage MCP servers from settings: dsh-plugin-setting-mcp.
Search GitHub issues and pull requests without leaving the chat: dsh-github-picker.
A keyboard command palette: dsh-spotlight. A pomodoro timer: dsh-pomodoro.

And a sample of the absurd, which is exactly what proves everything is a plugin: a desktop pet (dsh-desk-pet), a pixel whale that taps the glass and blinks the tab title when the agent is waiting for your approval (whale-on-desk), or a TikTok-style video feed in a floating window (dsh-plugin-video-player). If that can be built as a plugin, anything can.

And since adding plugins by hand is clunky, there’s already a plugin market inside the program itself (dsh-market): a searchable screen with one-click install and theme switching. All of this in the first week — a sign of how fast the community moved.

The four modes

DeepSeek Harness ships with four out-of-the-box configurations, each for a different job:

Standard: the full coding agent — file editing, terminal, search, skills. The everyday setup.
Code: the above plus a TypeScript SDK to orchestrate multi-step tasks in code.
Minimal: a two-tool agent (bash + editor), built for clean benchmarking.
Creator: for building your own plugins by chatting — you describe what you want the agent to be able to do, it asks questions and builds the plugin for you.

Full transparency: the trajectory log

DeepSeek Harness keeps a complete, ordered record of everything the model sees: system prompts, its reasoning, every tool call and result, and every bit of context injected. You can inspect it, pause it, fork it and replay it step by step.

And here’s the juicy bit: Anthropic and OpenAI deliberately hide that raw reasoning — out of fear it’ll be used to copy (distill) their models. DeepSeek does the opposite and shows all of it. For anyone who wants to understand why the agent did what it did — or tune its behavior by editing the system prompt — that’s a huge difference. All that instruction text steering the agent is exactly what context engineering works on.

The risk the hype skips

Everything being a community plugin has a B-side almost no article mentions, and it matters: installing a plugin runs third-party code on your machine, with your own permissions. That plugin can read your files, use your credentials and reach the network.

Heads-up:Watch out for one misconception: the agent’s tool approvals (that “allow this action?” popup) do not sandbox a plugin’s code. And being on a community “awesome” list — the curated resource lists GitHub is full of — is not a security review. Before installing a plugin you don’t know, read its source, and try it somewhere that doesn’t hold your keys or sensitive data.

DeepSeek Harness vs Claude Code and Codex

The comparison is unavoidable — plenty of people search straight for a Claude Code alternative or a Codex alternative — so here’s the honest version:

Claude Code / Codex

Polished, closed, ready-to-use tools. Plug in and go, but you only change what they let you change.

DeepSeek Harness

Less “finished product”, more platform to build your own agent. Everything is modifiable, but it asks for more tinkering (and care with plugins).

Put plainly: DeepSeek Harness is more open and powerful for anyone who wants to customize everything, but for that same reason it’s more developer-oriented than something you just want to work untouched. As a measure of how much harness design matters: Claude Code’s system prompt once ran to roughly 10,000 tokens (recently trimmed by 80%), while other harnesses run on barely 200. Those choices change performance and cost even when the underlying model is the same.

How to install and get started

Getting started is straightforward. You need Node.js, and one command launches the web UI (by default at http://127.0.0.1:3080):

npx @deepseek-ai/dsh web

You can also clone the repo and run it from source with pnpm:

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Since it’s free and open source, there’s no license to pay to try it; what can cost money is the model you connect over an API, depending on the provider.

Heads-up:Remember it’s in developer preview: the docs warn, in capital letters, that there will be compatibility-breaking changes. It’s an early build to experiment with, not (yet) a stable tool for production.

Why it blew up

Three reasons explain 177,000 stars in a week. One: the harness has become AI’s new battleground — models look more and more alike, so the difference is made by the layer that wraps them. Two: “everything is a plugin” thrilled developers, and within days a whole ecosystem appeared — hundreds of plugins, a market, community lists, desktop clients. Three: being free, open and from DeepSeek gives it obvious appeal against the closed tools from the big players.

It’s not the first open-source AI project to blow up out of nowhere. We recently explained PixelRAG, another case of a technical idea going viral in days. It’s the 2026 pattern: when something suddenly opens a new way of doing things, the community piles in.

FAQ

What is DeepSeek Harness?

It’s DeepSeek’s open-source coding agent: a software layer (a “harness”) that wraps an AI model and turns it into an assistant that can read files, run commands and work on its own through long tasks. Its distinctive trait is that everything — model, tools, interface, memory, even the agent loop — is a swappable plugin. It launched on August 13, 2026 under an MIT license.

What does “everything is a plugin” mean?

That every part of the system is a piece you can change, remove or replace from configuration without touching the core code: the model, the tools, memory, sessions, the interface, the theme and even the agent loop. Plugins install with dsh plugin add, and there’s even a market inside the program itself. Under the hood it’s made possible by a framework called Cordis.

Is DeepSeek Harness free?

Yes, it’s free and open source under an MIT license: there’s no license to pay to install or use it. What can cost money is the model you connect over an API, depending on the provider you choose.

Is DeepSeek Harness a Claude Code or Codex alternative?

Yes, it does the same job — turning a model into a coding agent — but with a different philosophy: instead of a closed, ready-to-use tool, it’s a platform where everything is a modifiable plugin. It’s more open and customizable, but also more geared to developers who want to tinker.

Is it safe to install DeepSeek Harness plugins?

With care. Installing a plugin runs third-party code on your machine with your permissions: it can read your files and use your credentials, and the agent’s tool approvals do not sandbox that code. Review the source before installing plugins you don’t know, and test them in an environment without sensitive data.

Sources

DeepSeek — Harness (official docs): deepseek.com — what it is, the plugin architecture, modes, install and the trajectory log.
DeepSeek Harness on GitHub: github.com/deepseek-ai — source code, MIT license, install (npm and from source with pnpm) and the developer-preview notice.
Awesome DSH Plugin (community list): github.com/awesome-dsh-plugin — plugin categories and real examples, install with dsh plugin add, and the security warning about third-party code.
The Register — on DeepSeek’s harness: theregister.com — why the harness is the new battleground, and comparison with Claude Code, Codex and others.