Deski lives on your own machine: chat, office work and coding that runs for hours. Behind it, one gateway to 220+ frontier models — images, avatars, video and an infinite canvas on a single key.
电商平台主图套图 · GPT-Image 2
2×2 奢华 3D 雕塑海报 · Seedream 4.0
乐高微缩城市杂志风 · Seedream 4.0
指定模特替换 · GPT-Image 2
生鲜食品销售场景套图 · Seedream 4.0
等距旅行海报 · Seedream 4.0
一只可爱的猫咪 · GPT-Image 2
促销大促海报 · GPT-Image 2
生成人物互动场景 · GPT-Image 2
科学概念主体海报 · Seedream 4.0
自定义描述换背景 · Seedream 4.0
小红书封面 · GPT-Image 2Not a chat box — a colleague who takes the job: reads your files, runs your commands, drives a browser, and keeps going for hours without babysitting. This is Deski — pick a task on the left or switch to Chat and watch it work.
Every session gets its own Git worktree, so refactors, bug fixes and test runs never overwrite each other. Switch back any time; merge when it's done.
Plan mode lists the steps before touching a file. Afterwards you get a readable diff: stage by hunk, commit, push, or open a Pull Request.
Ticket triage, alert sweeps, dependency bumps, build-failure digs — set a time and hand it over. Each run gets its own worktree; results land in your inbox for review.
Files are read and written on your disk, commands run in your shell, the repo stays local. Nothing gets uploaded to a container you can't see.
OpenAI, Anthropic and Gemini protocols side by side — just point base_url at the gateway. Account pools, automatic failover, unified billing, pay-as-you-go. Deski, Cursor, Dify and your own app all share one key.
from openai import OpenAI client = OpenAI( base_url="https://usapi.dflop.top/v1", api_key="sk-gpushare-····e7a1", ) r = client.chat.completions.create( model="gpt-5.6-luna", # or claude-opus-4-6 / gemini-3-pro / glm-5.3 messages=[{"role": "user", "content": "write a rate-limit middleware"}], stream=True, ) for chunk in r: print(chunk.choices[0].delta.content, end="")
Product shots, model swaps, posters, travel art — write a prompt or reuse a plaza template and get four images at once. Everything below was made on the platform.
Seedream 4.0Seedream 4.0Seedream 4.0
GPT-Image 2Seedream 4.0Seedream 4.0Seedream 4.0GPT-Image 2
Seedream 4.0
Seedream 4.0Seedream 4.0GPT-Image 2
Seedream 4.0
Seedream 4.0GPT-Image 2
Seedream 4.0
Seedream 4.0
GPT-Image 2
Seedream 4.0GPT-Image 2Seedream 4.0GPT-Image 2
GPT-Image 2
Seedream 4.0Upload a likeness, paste a short-video link to extract the script, clone your voice — a talking-head clip in minutes. Product pitches, lessons, company intros, no studio booking.
A front-facing photo or a short video is enough; real people go through the consent flow.
Paste a Douyin / Xiaohongshu link to pull the voice-over, or let a model rewrite it for your product.
Clone your voice, subtitles auto-aligned and burned in, export 9:16 / 16:9.
Input · one photoOutput · talking head · 1080p00:00
Text-to-video, image-to-video, reference rewrites, billed per second. Jobs run asynchronously — close the tab and come back; results are stored automatically, on domestic storage for domestic models so playback stays smooth.
References, prompts, generations, edits and video nodes all connect on one infinite canvas. Redo any step from that step; the whole chain stays visible.
Reference → prompt → generate → edit → video. Every node reruns; change an upstream node and regenerate downstream in one click.
Seedream for images, GPT-Image 2 for edits, Seedance for video — on the same canvas, no page switching.
Team members share the asset library and canvases; admins publish to the plaza or file into the knowledge base in one click.
Use them in the browser, @-mention them from Deski, or run the whole thing on your own servers with the enterprise edition.
Up to four models answer the same question side by side, streamed, with collapsible reasoning — pick the winner at a glance.
Open chat →AgentsPackage prompts, knowledge and tools into an agent; share it with the team or expose it as an API.
Create an agent →Knowledge base · cited answersDocuments become linked knowledge pages; an agent retrieves across them and pins every sentence to its source — or says it can't find it.
Upload documents →Speech · upscaling · subtitlesVoice cloning and dubbing, 4K video upscaling, automatic burned-in subtitles — the last mile of the pipeline.
See everything →Covers OpenAI, Anthropic, Google, ByteDance Doubao, Zhipu, DeepSeek, Qwen, MiniMax, Moonshot, xAI and other major vendors. New models are available within 24 hours of release.
Free credits on sign-up, no card, no minimum. Generate an API key in the console; Deski signs in with the same account.
