DeepSeek Harness Just Got a Web Agent

Author: Prompt Engineering · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek Harness Just Got a Web Agent · Channel: Prompt Engineering · 12:31 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Software engineering with AI, AI developer tools

Summary

The video introduces the DeepSeek Harness desktop app, describing it as a native shell for the existing harness and highlighting its plugin-based design, interface, and visibility into agent activity and token caching. It demonstrates extending the harness with Tiny Fish, an external web infrastructure service that provides search, fetch, web-agent, and browser capabilities through an MCP server. Examples show searching and fetching pages, logging into a site, filtering Hugging Face models, monitoring a release page, and reusing a saved login. The presenter also discusses costs, background runs, and how plugins make the harness extensible.

Key points

The desktop app runs the DeepSeek Harness process locally and displays its web interface in a native window.
The harness is built around plugins, which can be enabled, disabled, or configured through profile patch files.
The interface shows agent trajectories, tool calls, token usage, tool durations, and cache hit rate.
The harness offers standard, PTC, minimal, and creator modes.
Its built-in web search and fetch tools retrieve information but do not interact with logged-in websites or click page elements.
Tiny Fish connects to the harness through its MCP client plugin and provides search, fetch, web-agent, and cloud-browser capabilities.
Demonstrations include logging in to a test site and retrieving Hugging Face models that meet specified criteria.
Tiny Fish can monitor a page for changes and reuse a saved browser session without exposing the password to the agent.
The presenter says search and fetch are free, while two demonstrated agent runs each cost well under a dollar.

Chapters

Start	Chapter	Summary
00:00	DeepSeek Harness desktop app and plugin design	The presenter introduces the desktop app, its plugin-based architecture, and the interface features that show agent activity and efficiency.
03:10	Web access and Tiny Fish	The video explains the limits of the harness’s built-in web tools and introduces Tiny Fish’s web products.
06:47	Connecting Tiny Fish through MCP	The presenter shows how to configure Tiny Fish as an MCP server in the desktop profile and demonstrates search and fetch.
09:22	Web-agent tasks and monitoring	Examples cover filtering Hugging Face models, monitoring a release page, and reusing a saved login.
11:29	Costs, background runs, and wrap-up	The presenter discusses run costs and background execution before concluding that external plugins extend the harness’s capabilities.

Quotes

the app is just a host and the plugins are the product. — 00:31

they show you the cash hit rate, which none of the other harnesses actually shows you. — 02:07

you give it a goal in plain English, and it navigates, clicks, fill out forms, handles logins, and returns structured JSON. — 05:46

The built-in tools simply cannot do it at the moment. — 08:50

this extensibility is what makes the Deep Seek Harness itself extremely powerful. — 12:31

Concepts and entities

DeepSeek Harness, Tiny Fish, MCP, Marco, Hugging Face, Playwright, Claude Code, Codex

Full transcript (original language)

00:00 Deep just released desktop version of the Deep Seek harness, which is one of the fastest growing agentic coding harness in the open source space. In fact, I would argue it’s one of the best implementation, especially with their plugin system and their prompt caching. It’s one of the minimal but extremely powerful agentic coding harness. Now, if you have used Deep Seek harness before, you know that it’s a terminal app with a web interface. The desktop version doesn’t reinvent any of that. It’s a native shell around the

00:31 full harness web app. So, when you open it, it starts a local harness process in the background, picks a port, and loads the interface in a window. But, the most interesting part is how this is put together. A good mental model of this is that the app is just a host and the plugins are the product. Even the model routing, the web search, the terminal, the whole agentic loop is essentially a plugin package. Now, the beauty is that you guys can customize this. Each profile has one

01:03 patch file where you turn plugins on, off, or configure them, and the app has a plugin manager on top of that. Now, that design matters for what we are about to do because it means adding a new capability is not code change. It’s just a new file that you add to the plugin system. Now, before showing you a practical example of how you can extend the capabilities of the harness itself, let’s look at the UI because they do some really interesting things that I haven’t seen in other harnesses.

01:35 First, the chat interface is very similar to what you have seen before. However, with every agent turn, it also shows you all of the trajectories. This includes the context that you are providing, the user message, then all the steps the assistant is taking, including specific tool calls, how many tokens are used, what is the duration for every tool call, which is pretty amazing. This gives you visibility to the internal working of the agent itself. Uh this also shows you how many

02:07 tokens you have you consumed so far, and especially they show you the cash hit rate, which none of the other harnesses actually shows you. This gives you a really good idea of how efficient the inference side is. Also, for DeepSeek C Harness, this is one of the best harness when it comes to cash hit rate that I have tested. So, it really preserves your cash, and in return, you pay a lot less. It comes with four different modes. One is a standard mode, uh which works with

02:38 code files and information. Then there’s a PTC mode that includes all the standard mode capabilities, but it’s better suited for tasks that uh requires tool calls in batches, and then filter, organize, and deplicate. Then there’s a minimal mode, uh which is very similar to Pi with just a few different terminal tools, but you can build on top of it. And then there is a creator mode, which really gives you the capability to customize your whole agentic harness experience for your specific needs.

03:10 There’s a new video coming up on the creator mode because I think this is very close to what Claude Claude is trying to do with their mods feature. Now, in this video, I want to show you how to extend the capabilities of the harness itself. And for this, we’re going to be specifically looking at web access because it has become an integral part of any long-running agent. Out of the box, DeepSeek C Harness actually ships with two different tools for web. The first one is web search, which runs DeepSeek C’s on search inside

03:42 a model turn. The second is web fetch, which downloads a public page and gives the model its text. Basically, it can scrap any web page that you point it to. Now, both of these are reading tools. The fetch tool is deliberately anonymous, so it can send request, but it doesn’t send any cookies or credentials. And it doesn’t run the page like a browser would. So, it’s great for fetching information, but if you want your agent to take actions on a web browser, say things

04:14 like the tasks that are behind a login or it needs to click buttons, it will not be able to do it. There are some experimental browser use and computer use plugins in the works, but they are experimental and they are not installed. And you will need to be wiring up your own local browser to use them. Now, for an agent to be actually useful today, it needs to do a lot more than simply doing web search. Because a lot of generally useful stuff sits behind

04:44 logins and dashboards. So, the question becomes, how do you give this agent hands on the web without building that yourself? Especially for a very powerful agentic coding harness like Deep Seek Harness. This is where Tiny Fish comes in. I have partnered with them for this video to show you how to add new capabilities to the Deep Seek Harness. Tiny Fish is web infrastructure for agents. Think of it as a layer between your agent and the live web. In

05:15 this case, your agent decides what it needs, hands the web part to Tiny Fish, and gets a clean processed result back without any the messy HTML part. Tiny Fish gives you access to four different products in one API. So, they have web search, which gives you access to live web results. Then, you can use the fetch, which takes a URL, renders it in real browser, and hand backs that page as a clean text.

05:46 Now, the good part is that both of these are free on every plan at any volume. So, if you’re just interested in those two things, you can use this out of the box to the API. Then, there is a web agent. Essentially, you give it a goal in plain English, and it navigates, clicks, fill out forms, handles logins, and returns structured JSON. It runs on Tiny Fish’s own model called Marco, and you pay per step with the model included. And finally, browser, which gives you a

06:17 cloud browser. You can drive yourself with Playwright with antibot protection built into the browser engine and a vault for credentials. So, your agent can log into it without your passwords being exposed. Now, getting started is quick. You sign up without a card, and you get some credits in your own wallet to play with. Search and fetch are free. The wallet is only for agent and browser. Okay. So, let’s look at how you would extend the capabilities of the harness

06:47 itself with these new features. So, Tiny Fish runs an MCP server, and Deep Seek harness has an MCP client plugin. Now, in this case, in the desktop profile patch file, you will need to add the MCP client plugin. You will need to give it a server name, point it to Tiny Fish endpoint, and pass the API key as the header. Now, you don’t have to write the API key into the file. You can just read this off of an environment variable, which is a lot safer.

07:18 So, this way, you can share this config without leaking anything from your system. Next, we need to restart the app. We’ll open a new session, and the Tiny Fish tools are going to show up with everything else. And now, you have the ability to use all of those four capabilities inside the Deep Seek harness. Now, you can use Tiny Fish with other agentic coding systems like Cloud Code, Codex, Cursor, etc. There are way simple installation instructions that you can find. All right, so let me show you some of

07:49 these capabilities. So, first we’re going to start off with search. I’ll ask the agent to search for the DeepSeek-Coder soft release and show me the results exactly as they come back, not a summary. And you get structured results, which include things like titles, links, and snippets. These are live and for free, as I mentioned before. Okay, next is fetch. So, I’ll ask the agent to fetch the Hacker News front page through TinyFish and tell me the top story. And it’s actually able to do it almost

08:20 instantly. So, we are getting the clean text back. There are no menus or clutter, all right? And the good thing is that it’s very token efficient as well. Now, the thing that the default web access in DeepSeek-Coder harness cannot do is interacting with the websites. Let’s look at login. So, here I’m using a public test site that publishes its own demo username and password, and I’m asking the agent to log in, read what the page says, and log out again, and

08:50 give me the result as JSON. And on the right, you can actually watch the agent’s browser while it works. It opens the page, types the username, types the password, clicks login, read the message, and then logs out. That’s kind of exactly the task that we wanted to do. And the built-in tools simply cannot do it at the moment. Especially for repeated workflows, you want the agent to be able to interact with real web components.

09:22 Now, I wanted to try something harder, so I’m asking it to go to Hugging Face, filter to text generation models, then sort by trending, and give me the top five models that are 30 billion parameters or smaller. Now, the instructions were way specific, so I didn’t just say 30 billion or fewer. I told it to go down the list in order, read each model’s listed parameter count, and skip anything about 30 billion. And the thing is that the more explicit the prompt is going to be or the rules are, the more reliable the

09:53 results are going to come back. Okay, so this list has a couple of traps in it. One model has 29B in its name, but Hugging Face lists it at a 31 billion parameter, so it should be skipped. And in fact, it was. Another one is called 27B, but it’s really a 7 billion parameter model. It actually belongs in the list, and it found that, which is pretty great. Now, I also checked the results against Hugging Face’s own API, and the order matches exactly. And I’d highly

10:25 recommend doing this with any web agent. Now, if you look at it, you’re basically giving DeepSeek Harness computer use capabilities to interact with web interfaces, which is missing for open-source harnesses or open-source models. Okay, so two more things worth showing. Uh the first one is monitors. So, here I’m asking TinyFish to watch the DeepSeek Harness release page every 6 hours and tell me when a new release comes out. So, it takes a baseline right away, so it already knows the current

10:57 version, and from there on, it will only ping me when something meaningful changes. So, you get an agent that keeps watching the web for you, not just I mean, once for one specific run. The second one is saved login. So, instead of handing the agent a password, you log in once yourself in the TinyFish browser session, and save it as a profile. After that, the agent reuses that session, and uh it’s not going to ask you for password. Uh but the beauty is that it will never see your password.

11:29 Now, in terms of cost, search and fetch were free, and the login run and the hugging face run each cost well under a dollar in agent steps. Now, agents like these run were going to take a few minutes to complete. Uh so, I like to run them in the background, and the desktop app already shows that them as a background process. In some cases, the agent will wait for the runs to complete, so you might have to tell it for how often to check because if the agent keeps checking it out a few seconds, that might result in some token

12:00 cost just for waiting. Now, if you’re a Filecrawl user, Tiny Fish currently has an offer where if you send your latest Filecrawl invoice, you will get it matched in Tiny Fish credits. Which I think is pretty good because the search and fetch is already free for you. So, you can use those credits towards the agent runs. Now, this was a quick example of how you can extend the capabilities of an agentic harness like Deep Seek Harness with new powerful features using these

12:31 external plugins that you can bring in. And this extensibility is what makes the Deep Seek Harness itself extremely powerful. The beauty is that you can connect any model of your choice. So, I’ll highly recommend to check it out. Anyways, I hope you found this video useful. Thanks for watching, and as always, see you in the next one.