Wiring DeepSeek Harness into a remote workstation: 5 capabilities, 7 prompts, 2 things that break A developer documented a 16-minute walkthrough for extending the DeepSeek Harness (DSH) local agent into a remote workstation that generates images and video, serves output to phones, drives a browser and an Android device, and searches the web. The writeup flags two failure modes: published URLs are pinned to the machine's public IP and die silently when a dynamic address rotates, and model IDs must be entered as catalog IDs such as `deepseek-v4.1-flash` rather than the vendor's printed model name, which causes configs to save but calls to fail. DeepSeek Harness DSH ships as a local agent — it reads your files and runs commands on the machine it sits on. A 16-minute Chinese walkthrough takes it from that state to a machine that generates media, serves output to a phone, drives a browser and a real Android device, and searches the web. This is the build order, the seven prompts that were on screen in Chinese, with translations , and the two failure modes that will bite you before the interesting part: a published URL pinned to an IP you no longer have, and a model ID that does not match the vendor's name. Verified 2026-11-16. Written for developers who already run DSH; if you only use DeepSeek's own models for chat, you do not need any of this. AI disclosure: this post was written with AI assistance. The configuration steps reconstruct a third-party video; the gateway test at the end is my own run. Half the value of the walkthrough is knowing where you can stop. Gate each step on what your machine and network actually have. | Your situation | Steps that apply | Step to skip, and why | |---|---|---| | Laptop only, no GPU, no public IP | 1, 2 Agnes , 5 | 3 needs a public IP or a tunnel; 4 needs a phone for ADB | | Home server with a GPU, no public IP | 1, 2 ComfyUI , 4, 5 | 3 fails as written — put a tunnel in front instead | | Home server, GPU, dynamic public IP | all five | nothing — but expect to reissue links when the IP changes | | Android phone, no LAN interest | 1, 2, 4, 5 | 3 | The video runs the full path on a 4060 Ti 16 GB behind a home connection with a dynamic public IP, which is the only configuration where all five steps work as shown. Start here, because they cost more time than the setup itself. 1. The published URL is bound to an IP. Step 3 publishes generated files over HTTP using your public IP. The video's own agent output points out the consequence: the link is pinned to whatever address you had, so when a dynamic IP rotates, previously published links die quietly. The diagnostic the agent gave is worth keeping — check the port the publisher actually listens on 8899 in the demo , then forward that port on the router to the machine's LAN address. If you are behind carrier NAT with no public IP at all, this step does not degrade gracefully; it needs a tunnel instead, which the video does not cover. 2. Model IDs do not match vendor names. In step 1's add-model entry, what you type is the catalog ID. In this catalog that is deepseek-v4.1-flash or deepseek-v4-pro , not the model name printed on the vendor's page. Typing the vendor name is the most common silent failure: the config saves and the call fails. Neither of these is exotic. Both are the kind of thing you only find after something is already deployed, which is why they are worth reading before step 1. Install the DSH desktop client, sign in with a DeepSeek account, create a session, pick a working folder. Other models go in through the add-model entry; DeepSeek's own need no configuration. Remote access is a separate product, Agents Anywhere: install its client, sign in, then put its plugin inside DSH. The video has DSH install that plugin from a two-line instruction: 执行dsh plugin --profile desktop add agents-anywhere/dsh-bridge-next命令, 并参考https://(此处视频打码)完成插件安装 Run that dsh plugin add , then follow the reference doc to finish — the doc URL is blurred in the video, so it has to come from the Anywhere docs. Success looks like a remote-control item in DSH's left sidebar with the plugin running. Then pair the machine in Anywhere: "+", connect using the desktop program, and it appears in the device list with DSH sessions synced. Phones use the same product: connect phone, install the app, show the QR code, scan. Free, cloud: Agnes. Create a key and hand DSH the prompt plus the key. 参考文档帮我接入agnes的生图能力:https://(此处视频打码)/docs/agnes-image-25-flash,我的apikey是sk-... 然后帮我用同样的方式,再接入agnes的生视频能力,这是文档: https://www.agnes-ai.cn/zh-Hans/docs/agnes-video- https://www.agnes-ai.cn/zh-Hans/docs/agnes-video- ... Two operational notes from the video: the first call frequently returns 403 and a re-check by the agent usually clears it on the second attempt; and at recording time free video generation was queued into uselessness while image generation held. Treat this as a free image path. Local: ComfyUI. Install the desktop app, let it create the default instance, and the bundled templates cover image and video workflows MiniMax H3, LTX, Qwen, Flux, Z-Image . The prompt is one sentence and carries no configuration: 我电脑上已经安装好了ComfyUI了,请根据我电脑的配置选择合适的模型,调用ComfyUI来生图和生视频 ComfyUI is installed; choose models for this machine and call it to generate images and video. The agent reads the local CLI, inspects the hardware, proposes models — Z-Image for stills, MiniMax H3 for video on the reference build — and runs a test. Remote sessions render generated files as local paths, so on a phone or a remote laptop the images simply do not appear. The fix is a plugin DSH writes from a prompt that is worth reading as a specification: 我现在在用Agent Anywhere这样的远程连接DeepSeekHarness的工具。但是这个工具有一个问题:如果我做了一些生图、生视频或者生成文件,因为它是远程连接,默认是按照本地路径来生成的,所以是打不开的。 所以我现在想做一个插件,具体设想如下: - 触发机制:在我生成图或者视频的时候,它会去调用这个插件,把文件部署成一个HTTP服务。 - 网络配置:可以给它配置一个公网IP,或者让它能自动读取我电脑所在的公网IP。因为我的电脑是有公网IP的,但是它是家庭宽带的动态公网IP。 - 给插件的配置增加单独的一个配置页 Problem first, then intent, then three constraints: hook generation and serve over HTTP, accept or discover a public IP, and ship a settings page. After that, one phrase in your generation prompts — publish with the file-publishing plugin — is enough to give every output a public URL. The same mechanism carries the HTML reports in steps 4 and 5. Browser. egolite is a browser built for agents. Two commands, not prompts: where ego-browser npx skills add citrolabs/ego-lite The first confirms the install, the second installs the driving skill. The test pulls data from a video site, narrates each step, reports the sample count, and produces an HTML report published through step 3. Phone. Android ADB, explicitly wireless: 先让我的DeepSeek Harness具备接入操作真实手机的能力,使用安卓的ADB就可以了,而且支持无线ADB。 Workbench. One more prompt asks for a management surface over both: 围绕刚才接入的这两个能力,再给它做一个工作台页面,可以看到设备的执行过程以及监控执行任务的过程,并且将这一套能力和这个工作台做成一个DeepSeek Harness的插件。 The interesting part is not the UI but two behaviours: a human can take over mid-task and hand control back, and each task ends with a process report and a result report. Unattended automation is fine right up to the first task that gets stuck; the takeover switch is what makes it usable repeatedly. DSH has web search; the video layers AnySearch on top for 1,000 free calls a day, roughly 40 vertical data sources, and better results. Register, create a key, run the install instruction inside DSH, paste the key, and trigger it by name in the prompt — search with AnySearch. I wanted to check the one entry that reaches every non-DeepSeek model. I used an OpenAI-compatible gateway — base URL https://api.crazyrouter.com/v1 — configured in DSH's add-model entry. The test is a constrained optimization with a checkable answer: three shelf types A: 2 m², 80 items; B: 1.5 m², 50; C: 1 m², 30 , 400 m² available, A equal to twice B, at least 9,000 items stored, minimize floor area. deepseek-v4.1-flash returned A=86, B=43, C=0 for 236.5 m² and 9,030 items, with a full derivation, plus the observation that requiring all three types moves the optimum to A=84, B=42, C=6 at 237 m². I brute-forced it independently: same optimum, same totals. Usage is the part that matters operationally. Total 3,381 tokens, of which 2,643 were reasoning tokens, in 24.5 seconds. For a per-token budget that split is the cost driver, and it is the reason this class of model belongs in multi-step work with a verifiable answer rather than in rewriting or formatting. Verify the current IDs in the Crazyrouter model directory https://crazyrouter.com/en/models before writing one into a config; if you are checking billing assumptions too, the model list and pricing page https://crazyrouter.com/en/pricing?utm source=devto&utm medium=article&utm campaign=offsite article&utm content=cr cbad0bf9584c mid text is the place to confirm them. ego-browser resolves, the skill is added. The prompts here share one structure, and it is not "be detailed". Each states a situation, an outcome, and a constraint — my connection has a public but dynamic IP; publish generated files as HTTP; ask me when blocked. Nothing explains implementation, which is exactly why the agent can act. If you copy one thing from this walkthrough, copy the failure modes: the IP-pinned URL and the catalog-vs-vendor model ID. Those are the two that will cost you an evening.