Загрузка видео...
Не удалось загрузить видео
DeepSeek GUI just dropped and it’s exactly what power users needed. A clean, local-first desktop agent workspace for DeepSeek with: • Code Mode - real file ops, planning, reviews, approvals & project context • Write Mode - excellent Markdown editor with smart AI assistance • Kun runtime - insane... show more
150,064 просмотров • 4 месяцев назад •via X (Twitter)
Комментарии: 33
Why doesn't the official DeepSeek website mention anything about a GUI? It looks like a third-party app. Be careful, it could be malware unless DeepSeek itself confirms it.
the part that interests me: "cache-first design, high hit rates." at deepseek's pricing ($0.87/M output tokens) plus aggressive caching, the cost of running a 24/7 background agent just went from "need to justify the budget" to "costs less than your spotify subscription." the economics of always-on agents just crossed a threshold.
The most interesting part isn't the model, it's the workflow. We're moving from AI as a chatbot to AI as a workspace where coding, writing, planning, and automation happen together. If the experience is smooth and the token efficiency is as good as advertised, that's a bigger deal than most benchmark improvements.
Local-first desktop agent workspace for DeepSeek is the right call — sending every context window to a remote API adds latency, cost, and data exposure that most power users don't actually want.
Why does it look exactly like claude code🤔
@Grok - is this an official DeepSeek software?
There is no version for Linux :/
local-first, cross-platform, cache-hit rates above 85%, scheduled background agents, and it costs two cents a session. we're cooked (in the best way). the bar for "what a serious AI workspace looks like" just quietly moved and most people are still debating which cloud tool to subscribe to.
Is it just me or does this look exactly like Claude OS desktop app lol
Deepseek official website makes no mention. Is this GUI website legitimate?
It maximizes token conservation, ensuring that a single "hello" message doesn't cost you $1. This is a very clever and win-win business model.
I have no idea why it developed a coding harness from scratch, given that Codex and OpenCode is powerful and providing SDKs.
Codex
Where is linux
So clean and fantastic
Deep seek is just getting fireee 🔥
Useful setup for faster and more private AI work.
The approval step is underrated. Most agent frameworks just execute and hope for the best. A review gate before file ops saves so many headaches.
The Kun runtime cache-first design is the part I want to see benchmarked against Claude Code. Anyone actually tested this yet?
love seeing desktop agent workspaces take off. the thing single-agent ones still miss: running a few at once without them stepping on each other's files. been building for exactly that, each agent in its own worktree. here's the setup:
ish
🤔
Cross platform Windows and Mac only 😂
it is not official from deepseek team, watch out! could be dangerous
DMing you! 💬
@NexEcosystem China has released a new model, the Nex-N2-Pro, which is currently ranked number one in China. Its benchmark score exceeds GPT 5.5. You can check it out.
缺一个多模态就完美了，能够满足大部分人的使用需求
为什么不使用Reasonix？
Is it open source?
Bu tür bir arayüz, projeleri yönetmeyi çok kolaylaştırıyor. Kodlama sürecini hızlandırmak için mükemmel bir çözüm!
@MrAhmadAwais hey, have you compared command code harness vs deepseeks native harness (if that's what this is)
This looks solid. The Kun runtime's cache-first approach with Code Mode file ops and project context stands out for efficiency. Local-first on macOS/Windows is a good move for power users.
where can I find this?
![HERMES AGENT HAS FEATURES THAT 90% OF USERS NEVER TOUCH. /BACKGROUND. /STEER. /GOAL. SUB-AGENTS. MODEL SWITCHING. CRON JOBS. 7 LEVELS OF HERMES. ALL EXPLAINED.👇 LEVEL 1 — ONE-SHOT PROMPTS you type a prompt. agent responds. done. this is using Hermes as a chatbot. it works but you leave 90% of the value untouched. LEVEL 2 — MEMORY + SOUL.MD the agent remembers you across sessions. you wrote a SOUL.md with identity, voice, restrictions. Hermes tailors every answer to your context. two people asking the same question get different answers because it knows them differently. LEVEL 3 — COMMANDS THAT MULTIPLY OUTPUT /background → fires a task in the background. your main session stays free. result appears as a panel when done. /steer → injects a message into the current run without interrupting the agent. redirects the work mid-execution. /queue → queues a follow-up for after the current task finishes. /model → switches models mid-session. start with Sonnet for planning. switch to DeepSeek for execution. switch to Opus for review. configure default behavior: display: busy_input_mode: steer # or queue, or interrupt LEVEL 4 — SKILLS + RIGHT MODEL PER SKILL every installed skill becomes a slash command: /competitive-research analyze my top 3 competitors /social-media-copy write 5 LinkedIn posts /code-review check this PR for security issues the unlock: assign a specific model per skill. research skill → GPT-5.5 (cheap, high volume). code review skill → Fable 5 (best at code). content skill → Sonnet (best at writing). you stop paying premium prices for tasks that don't need premium models. LEVEL 5 — MCPs (CONNECT YOUR WORLD) plug Hermes into the tools you already use: Gmail, Calendar, Notion, Slack, ClickUp, Granola. the agent reads your emails, checks your calendar, pulls from your project management tool, and answers questions using YOUR actual data. "what were the numbers this week?" Hermes checks ClickUp, cross-references with your goals in memory, gives a contextualized answer. caveat from the docs: keep MCPs minimal. every MCP adds tools to the context window. 15 MCPs with 10 tools each = 150 tool schemas the model reads every turn. install what you use. disable what you don't. LEVEL 6 — SUB-AGENTS + PARALLEL EXECUTION delegate_task spawns sub-agents with isolated context. batch mode runs children in parallel: delegate_task(tasks=[ {goal: "research best industries for websites", model: "deepseek-v4-flash"}, {goal: "critique the research and find gaps", model: "gpt-5.5"}, ]) max_concurrent_children: 3 (default). each child gets own terminal session and toolset. parent receives summaries when children finish. roles: → leaf (default): cannot re-delegate → orchestrator: can spawn its own workers (bounded by delegation.max_spawn_depth) LEVEL 7 — ASYNC OPERATIONS this is where the agent works without you. /goal — persistent objective across turns. judge model checks after every turn. runs until done or budget hit (default 20 turns). cron jobs — scheduled tasks on repeat. morning briefs, competitor scans, weekly reviews. wakeAgent gates for zero-token monitoring. /background — parallel tasks in current session. results delivered when ready. your main conversation never stops. the shift: from "I ask, it responds" to "it works, I review." WHERE ARE YOU? 1 → chatbot 2 → it knows you 3 → parallel commands 4 → right model per task 5 → connected to your tools 6 → sub-agents working together 7 → runs without you most people are at 2-3. the jump from 4 to 7 is where the real time savings compound. this is just the first 7. beyond this: multi-profile architecture, self-improving knowledge bases, voice mode, kanban orchestration, webhook automation, profile distributions, browser control, IDE integration, and more. comment LEVEL and I'll send you the advanced levels with setup for each. Full guide how to build 3-agent research department with Hermes 👇](https://image.24vids.com/tw-2067607662582284458/amplify_video_thumb/2067603833216221184/img/iYKSGkZ71CVRuf4D.jpg)