One of the fastest-moving open-source coding harnesses just moved onto the desktop: DeepSeek Harness now ships as a native app you install, point at a folder, and start working with. It sounds like a small convenience, no terminal and no Node setup, but for long-running agent work, removing setup friction puts the real product, the plugin system, in front of everyone. DeepSeek Harness is turning the coding helper into a daily tool.
A host app whose product is plugins
Start with the right mental frame: the app is only a host, the plugins are the product. Model routing, web search, terminal access and the whole agent loop live as a plugin bundle; each profile flips plugins through a single patch file, with a plugin manager on top. According to the Github repository, the project is MIT-licensed, built on the Cordis runtime, and still in developer preview, so breaking changes are on the table. The DeepSeek official page presents the same story as a worldwide public preview with composable plugins.
The desktop release does not rewrite the web interface; it wraps the full web app in a native shell. On launch, a local harness process starts in the background, takes a free port, and loads the interface into a window. The Github docs show the same experience opening from the command line with a single command served over loopback. The design keeps the shell thin and puts all the weight into the plugin runtime; adding a capability is not a code change, it is a new file dropped into the folder.
What caught my eye is the trace view next to the chat. Every agent turn lays out the context, the user message, each step taken, tool calls, tokens spent and per-call duration, with total consumption in the open. What sounds like cash in the audio is really the cache-hit rate : hits translate directly into lower inference bills, and in the presenter's tests this harness comes out notably thrifty. Showing the internals this openly is a trust-winning choice for long jobs.
Four modes serve different jobs: standard for code and knowledge work, PTC for batched tool calls that get filtered and organized, a minimal Pi-like starting point with few tools, and a creator mode for shaping the whole experience. The creator-mode example on the DeepSeek page makes the claim concrete: a floating timer plugin dictated through chat gets built and verified in about five minutes. The presenter notes this lands close to what Claude Code does with mods; the name heard as Cloud in the audio is actually Claude Code .
Reading is not enough: the web behind logins
The two built-in web tools only read: one runs web search inside a model turn, the other downloads a public page as text. The fetcher deliberately stays anonymous, with no cookies, no credentials, and no running pages the way a browser would. So anything behind a login, any button clicks or form fills, stay out of reach for these two. Experimental browser-use and computer-use plugins exist but ship uninstalled; you wire up your own local browser. Given how much daily work sits behind logins and dashboards, the gap sits exactly here.
The gap gets filled by a layer between the agent and the live web: TinyFish . One API key and one wallet back four products; live search and page fetch stay free on every plan at any volume, while the goal-driven web agent and the cloud browser bill per step. On the TinyFish side, search hits a real browser-rendered live index and returns structured JSON; a password vault plus saved sessions lets the agent finish logged-in work without ever seeing the password, and step-by-step run recordings keep everything auditable. The model heard as Marco is really Mako : per the TinyFish announcement of July 22, 2026, a narrow specialist trained on real logged-in, multi-step production work, now the default engine behind the web agent. Single-mindedness lets it spend a fraction of the tokens frontier generalists burn, so latency and cost fall together.
Single-file wiring and the live exam
Wiring is as plain as a handshake between an MCP server and a client plugin: the desktop profile's patch file takes a server name, an endpoint, and the API key carried in a header, read from an environment variable instead of pasted into the file, so the config can be shared without leaking secrets. Restart, open a fresh session, and the TinyFish tools land among the rest; the same pattern extends to code tools like Claude Code, Codex and Cursor. According to the Microsoft Learn guide, this pattern states a general principle: plugging an MCP server into an agent turns it into a declarative tool in four steps. This is how capability gets added without writing code.
Live trials confirm the layer. Search takes a soft-release query verbatim and returns structured title-link-snippet triples; fetch turns a news front page into clean text with no menu residue, token-thrifty. The login trial runs against a public demo site: the agent types the credentials, reads the message, logs out, and answers as JSON while the browser stays watchable on the side. The hard trial sorts a HuggingFace text-generation list by trending and collects the top five at 30B parameters or below, skipping the one named 29B but listed at 31B, keeping the one named 27B that is really 7B, with results matching the HuggingFace API exactly. The HuggingFace listing behind such traps, Qwen3-30B-A3B, shows 30.5B total parameters with only 3.3B active across 8 of 128 experts. The rule is plain: the more explicit the instruction, the more reliable the result.
Two final pieces make the work continuous: the monitor and the vaulted login. The monitor watches the release page every six hours; it takes a baseline first, then pings only on meaningful change. On the vault side, a session gets opened by hand once and stored as a profile, and the agent reuses it afterward without asking for, or seeing, the password. Independent examples of the same class exist: Scrapfly docs describe a cloud browser driven over CDP with a 21-command antibot domain, while Gologin's side keeps persistent sessions holding cookies and authentication state so one login keeps working for months. Costs stay plain: search and fetch are free, and both agent runs land well under a single unit of currency; long runs go to the background, with only polling frequency to watch, since a chatty agent burns tokens just by waiting. The service heard as Filecrawl is really Firecrawl, currently matching invoice credit for a limited time.
| Design | Connection |
|---|---|
| Host app | Product is plugins |
| Search and fetch | Free, unlimited |
| Vault and monitor | Passwordless continuity |
Key moments
AI commentary
"A design that treats plugins as the product, paired with a web layer that can actually log in, lets open-source coding agents speak the language of daily work for the first time. Without reaching for hype: this is one of the rare extensions that grows capability instead of chrome."
AI assessment
The strongest objection: browser agents are brittle; login flows, antibot walls and interface redesigns can break yesterday's working instruction today. Narrow specialists like Mako run cheaper and faster than general giants, but on a site's odd trap, general breadth can save the day. That is why the presenter's advice to cross-check against the HuggingFace API is not decoration, it is the seatbelt for this class of work.
Gaps show too: failure cases and broken runs never appear; beyond the free tier, pricing stays a per-step sentence with no totaled bill for long runs. The Firecrawl offer is flagged as limited, but its end conditions stay vague. The monitor's false-alert rate, and whether a six-hour cadence serves urgent work, also go unanswered.
The commercial angle needs stating: this is a partnered video, produced with the TinyFish team. That does not make the claims false, but it selects the frame: rival layers and free open-source options never enter the shot. On security, the vault design points the right way; still, opening logged-in portals to an agent demands least privilege: split read and write, separate profiles, rotated secrets.
The practical takeaway stays plain: connect the free search and fetch over MCP first and put them on daily work, then grow into vaulted logins and monitors for the heavy jobs. Running long work in the background, writing the polling interval into the instruction, and keeping keys in environment variables cuts both cost and risk. Without measurable targets, how many logged-in jobs a week and how many correct results, the layer's real value stays unmeasured; track those two numbers in week one.
Sources
9 links; 1 of them also cited by 1 other story. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
deepseek · harness · tinyfish · mako · mcp · plugins