DeepSeek Harness Just Got a Web Agent

0:00DeepSeek Harness (dsh) — one of the fastest-growing open-source agentic coding harnesses — just shipped a desktop version, and the interesting part isn't the native window. It's the plugin architecture underneath, which the video uses to show how to bolt a full web agent onto the harness without touching its source. Fair warning up front: the second half of the video is a sponsored walkthrough for TinyFish, a web-infrastructure-for-agents company. The genuinely useful, non-sponsored substance is the first half — the harness design — and the general technique (MCP plugin) it demonstrates.

The desktop version

0:21DeepSeek Harness was already a terminal app with a web interface. The desktop release doesn't reinvent it: it's "a native shell around the full harness web app." Opening it starts a local harness process in the background, picks a port, and loads the interface in a window. The video's assessment: "one of the best implementations" in the open-source space, especially the plugin system and prompt caching — "minimal but extremely powerful."

"Everything is a plugin"

0:41The mental model: "the app is just a host, and the plugins are the product." Model routing, web search, the terminal, the whole agentic loop — each is a plugin package. This is the Cordis architecture (the harness's "everything is a plugin" runtime): every capability, from models to sandboxes to scheduling, is a plugin that can be selected, swapped, or extended in configuration without changing the harness source.

"Adding a new capability is not a code change. It's just a new file that you add to the plugin system."

1:00In practice: each profile has one patch file where you turn plugins on, off, or configure them, plus a plugin manager on top.

The UI: trajectories and cache hits

1:23The chat surface looks familiar, but each agent turn exposes the full trajectory: the context you supplied, the user message, every assistant step with its specific tool calls, the token count and duration per tool call. "This gives you visibility into the internal working of the agent itself."

2:00The standout: it shows your cache hit rate — "which none of the other harnesses actually shows you." That's a direct window into inference efficiency, and the video's claim is that DeepSeek Harness is "one of the best when it comes to cache hit rate… so it really preserves your cache, and you pay a lot less."

The four modes

2:34The harness ships with four operating modes:

• Standard — code files and information.
• Batch/parallel-tool mode (the transcript's acronym is garbled) — standard capabilities plus batched tool calls, then filter, organize, and deduplicate.
• Minimal — "very similar to Pi," just a few terminal tools you can build on.
• 3:02Creator — customize the whole harness experience, which the video compares to Claude Code's mods feature (a separate video is promised).

Built-in web access (read-only)

3:32Out of the box, the harness has two web tools, both readers. Web search runs DeepSeek's own search inside a model turn; web fetch downloads a public page and returns its text. The fetch tool is deliberately anonymous — it sends no cookies or credentials and doesn't execute the page like a browser would.

4:34The gap is exactly where real work lives: "a lot of generally useful stuff sits behind logins and dashboards." Fetch can read a page but can't click a button or get past a login. So the question becomes how to give the agent hands on the web — which is where the extension part begins.

Extending it: MCP + TinyFish

4:58The sponsor enters here. TinyFish is "a layer between your agent and the live web": your agent decides what it needs, hands the web part to TinyFish, and gets back clean processed results without the messy HTML. Four products in one API: search (live results), fetch (renders in a real browser, returns clean text), a web agent (a plain-English goal becomes navigation, clicks, form-filling, and logins, returning structured JSON, running on TinyFish's own Mako model), and a browser (a cloud browser you drive with Playwright, with anti-bot protection and a credential vault).

6:43The integration is the part worth generalizing: TinyFish runs an MCP server, and DeepSeek Harness has an MCP client plugin. You add the client plugin to the desktop profile's patch file, point it at the endpoint, and pass the API key via an environment variable rather than hard-coding it ("you can share this config without leaking anything from your system"). Restart, open a new session, and the four capabilities show up alongside everything else. The same pattern — an MCP server wrapping some external capability, wired in through a plugin file — is the reusable technique, regardless of vendor.

What the web agent can do

8:30The demos show the delta over the built-in tools. On a public test site, the agent logs in, reads the page, logs out, and returns JSON — you can watch the browser on the right as it types the username, types the password, clicks login, and reads the message. "The built-in tools simply cannot do it at the moment."

9:22The hardest test is genuinely instructive: go to Hugging Face, filter to text-generation models, sort by trending, and list the top five models ≤30B parameters. The prompt was made deliberately explicit — go down the list in order, read each model's listed parameter count, skip anything over 30B. The traps: one model has "29B" in its name but HF lists it at 31B (must skip); another is called "27B" but is really a 7B model (must include). The agent got both right, and the video verified the ordering against Hugging Face's own API — which it recommends doing with any web agent. The meta-lesson: "the more explicit the prompt or rules are, the more reliable the results."

10:41Two more features round it out: monitors (watch a release page every 6 hours and ping only on a meaningful change — an agent that keeps watching the web, not a one-shot run) and saved logins (log in once yourself, save it as a profile, and the agent reuses the session without ever seeing your password).

11:29Cost-wise: search and fetch are free; the login run and the Hugging Face run each came in "well under a dollar" in agent steps. One practical note — run these in the background and tell the agent how often to poll, or it'll burn tokens re-checking every few seconds.

Claims checked

Sources & further reading