Martin Tang

@aigclink MCP 把时间线开放给 Agent 很关键。但真正决定能不能用的，是每一步能否预览、撤销和回滚。

2026 年，剪辑器正在被 agent 接管——注意，是"接管"，不是"加个 AI 按钮"：剪辑正在从"你用鼠标拖时间线"，变成"你和 agent 同编一条时间线"。一个多月里，我连着看到了三条不同的路径。 它们的共同暗号都是：MCP协议——剪辑器把自己的时间线开放给 agent 操作的通用接口。谁把它接上了、接到什么程度，决定了这三家站在哪一格： 路径一：给现有编辑器外挂接口 —— ChatCut 把一个闭源 SaaS 剪辑器整个开成托管 MCP，你用自然语言让 Codex 帮你导素材、改时间线、加字幕、导出。最轻，但需要 ChatCut 账户，仓库也没 license。适合"我不换编辑器，只想给它接上 agent"。 路径二：从零重写成 agent-native —— OpenCut 68000 star 的开源 CapCut 平替，正在从头重写：Editor API、MCP server、headless 批量渲染、Rust core 一套代码通吃三端，全写进了路线图。最彻底，但——这些还没上线，今天能用的是 classic 版，期待它完整版早日上架，要是能直接支持聊天剪辑就完美了。 路径三：开源原生 + MCP 已上线 —— Palmier Pro Swift 原生、对标 Premiere、GPLv3、YC S24。MCP server 今天就能用（一行命令接 Claude Code/Codex/Cursor），时间线内置 Seedance / Kling / Nano Banana Pro 生成，还有 in-app agent——人和 agent 真的在同一条时间线上一起剪。走得最完整，代价是只支持 macOS 26 + Apple Silicon。 把三条路径叠起来，能看出这波"接管"到底在变什么： 1、接口层：MCP 正在成为剪辑器对 agent 的"通用语"——外挂也好、重写也好，终点都是把时间线变成 agent 可调的对象。 2、交互层：从"GUI 里点选"到"自然语言指挥 + 人机共编"。鼠标没消失，但它旁边多了一个能听懂"把这段的废话剪掉、配个背景音乐、导 1080p"的搭档。 3、商业层：不约而同走 open-core——编辑器和接口开源做流量，把生成算力留成闭源订阅（Palmier 最典型：不卖编辑器，卖生成）。 再往大了看，剪辑只是 AI 视频整条链里被 agent 化的一环：看懂视频 → 剪辑 → 生成 → 世界模型，每一环都在从"人肉工序"变成"agent 可调工序"，剪辑是这半年动得最快的那一环。 但别急着喊"剪辑师失业"——三家都还很早： ChatCut 三百多 star、无 license；OpenCut 的 agent 能力还停在路线图；Palmier 只跑在最新的 macOS 上。"接管"现在是方向，不是现状。 所以 2026 剩下的问题，已经不是"剪辑器会不会被 agent 接管"，而是——谁的 MCP 接口会先变成那个标准。剪辑器的下一个高频用户，可能真的不再只是人而是MCP协议。 #AI剪辑 #MCP #Agent #ChatCut #OpenCut #PalmierPro

@hnshah Seat pricing weakens, but usage pricing can punish customers for succeeding. The better unit is probably closer to completed work than headcount or tokens.

@zjp1997720 这套授权里最有价值的不是六个后台任务，是主 Agent 还负责集成和验收。并行很快，跑偏也会更快。

用了这个skill的Codex真是的又猛又省～ 主控模型用Sol xhigh Fast蹬了一天了，才花了20%的周额度 搭配以下系统提示词使用Codex会更主动： ## Codex 后台模型路由授权 - 用户长期授权 Codex 在复杂、可并行任务中自动使用 `$codex-model-routing-team` 创建独立后台任务，并为其指定模型与推理强度；派遣前用一条简短通知说明数量、模型、强度和职责，无需再次确认。 - 主 Agent 保持当前模型，负责规划、文件所有权、集成、验证和最终交付。 - 同时运行最多 6 个后台任务；单个根任务累计最多创建 8 个。后台任务不得再创建任何后台任务或子 Agent。 - 后台任务禁止使用 Ultra；Terra 默认不参与路由。无法使用 Codex App 后台任务接口时，主 Agent 本地完成，禁止回退到 MultiAgentV2 `spawn_agent` 冒充模型路由。 - 简单问答、状态查询、单文件小改、强顺序任务以及发布、发送、付款、删除、账户或生产操作不自动派遣。

@Sanjil @GoDrift_ai @nikhilkr The closed beta and customer proximity are more useful than the Product Hunt rank. What changed in the product because of those first three months?

It's been 1 year of building @GoDrift_ai. When we began, my co-founder @nikhilkr had unwavering conviction that this was the idea worth betting on. We'd already built and sold our last startup together, but we knew this journey would be on an entirely different scale. 12 months later, here's everything we've accomplished: >developer first bet: developer signups from 45+ countries now >ranked 5th on soft launch on Product Hunt - competing against Gemini and Anthropic >spent first 3 months in closed beta with constant feedback loop - just to form tunnel vision into product strategy. >shifted between 3 countries and took risks to get into the largest market faster (avoiding comfort/slow built) >we had a contrarian opinion on how the robotics software stack should be built and scaled: like our focus on scaling evaluation >>> data. We were absolutely right from day one. Decision of being closer to your customer worked great. >Got featured by Tech in Asia and Fortune :) humbled, and only motivated to build for a better future! I will keep sharing updates here, follow along :)

@BenBajarin Financial modeling is a good stress test because a plausible spreadsheet can still be wrong. Will the benchmark score traceability and formula repair, not just the final number?

For our AI model knowledge work benchmark we will keep adding knowledge work use cases, but financial modeling (forecasting, market sizing, financial analysis, etc.) is a pretty complex and token consuming task so a good one to test. csbench.com

@JasonrShuman A free application portal can copy the visible flow. The $100M question is who owns compliance updates, integrations, and support after launch.

One startup has raised $100M to scale their Application Portal for Credit Unions. Another startup, Fuse, just launched a Free AI Generated Application Portal for Credit Unions that is on brand and ready in minutes. If this isn't going to disrupt the banks and credit union market, I don't know what will.

@AYi_AInotes 通用工具被做成一个 tab 不可怕。可怕的是你的产品只负责把同一个 tab 换个皮。

我感觉OpenAI的野心已经很明显了，他们要做AGI时代的苹果，逐渐在构建自己得苹果生态了，那些做通用工具的创业公司真的要小心了，我预感OpenAI 会把你的品类变成一个 tab，想活下来，真得往更深的垂直工作流里钻。 这两天发现很多人在说 OpenAI 砍了 Codex，我研究了下，发现真相刚好反过来：不是 ChatGPT 吞了 Codex，是 Codex 赢了 ，说白了它只是钻进了 ChatGPT 这个流量更大的壳子里。 四层反直觉给大家讲清楚： 1️⃣第一层：强者为什么要装弱者？ 官方说 "简化体验、减少碎片化"， 没错，但这不是重点，真实情况：Codex 才是增长爆炸的干活引擎，将近800 万 + 周活，全公司包括非工程师每周都在用，能连续自主执行 7 小时以上。 也就是说ChatGPT 是品牌入口，Codex 是能力内核， 用大品牌给硬核能力插翅膀，比硬核能力自己获客快 10 倍。 2️⃣第二层：产品哲学彻底转向了 以前 ChatGPT 是 Q&A 神器 —— 你问，它答， Codex 证明了真正值钱的是 agent loop：规划→调用工具→执行→交付成品。 所以现在拆成三个模式：Chat（聊天）、Work（通用知识工作）、Codex（深度编码）， 不是加了几个 tab，是整个产品从 "回答问题" 升级成了 "完成工作"。 3️⃣第三层：最反直觉的判断 Codex 负责人 Andrew Ambrosino 反复说：现在软件构建太便宜了，便宜到 OpenAI 内部有 90 个团队同时在原型同一个东西。 但其实真正的瓶颈已经不再是 "能不能做出来"，而是 "该不该做" 和 "做成什么样"。 品味（taste）变成了稀缺资源，那么当人人都有执行队的时候，指挥官才值钱。 4️⃣第四层：更大的棋是 AI OS ChatGPT + Codex + Atlas 浏览器 + Skills 生态 + 未来硬件，目标很明确 ——AI 时代的超级 App，或者说新操作系统。 你打开一个应用，指挥一堆 Agent 干活，不用跳来跳去，像 Apple 生态一样把用户锁死， 很快模型公司的战争就要结束了，但平台公司的战争才刚开始。 最后给三类人一句话： 👉 普通用户：以后别光聊天了，试试 Work 模式，这可能是你离 收获一个AI 员工最近的一次 👉 开发者：保留 Codex 图标、设为默认视图，功能没少，入口变了而已 👉 创业者：做通用工具的真的要小心了，OpenAI 正在把你的品类变成一个 tab，想活下来，得往更深的垂直工作流里钻。

@ShubhAgrawal26 The Obsidian vault on the Mac mini is a great example of why remote access becomes part of the AI workflow. Did the setup preserve permissions or expose the whole vault?

something really basic I did today that made me feel like - "damn! AGI is almost here and technology has gone so far" I was building some content and growth infra which had an elaborate directory of thousands of LinkedIn posts, tweets, newsletters, and internal docs inside an Obsidian vault on my Mac mini. I then had to visit a hospital to meet a friend for his surgery, while in the hospital, I had my macbook with me. I realized I needed this obsidian vault's data to connect with ahref's, google search console and promptwatch - check our AI visibility and self create structures of blogs for AEO from existing past content, by identifying whatever is missing or noot ranking in our sitemap. but this vault was on my mac mini at home , which is also where I had the claude code instance running. I was able to ask claude code via remote control using my phone -> make a copy of the vault and paste it from my local mac mini folder to my icloud -> opened this vault on my macbook and complete the entire task. insane!! we're truly living in the future API's and MCP's fetch real time metrics and data from SaaS tools markdown files that seld update with relevant company context Agents that can control and take actions on your computer while you run them from ur phone complete analysis to great results end to end. beautiful

@thenarrator Distribution matters, but trust is not transferable forever. Robinhood still has to make the new product feel native to funded-account users.

the thing people underrate this cycle is how much of the winning is just distribution now the tech works, the rails are fast enough, the products are good enough what separates winners is who can put something in front of people who already trust them. that’s why robinhood and coinbase matter more than any startup, they sit on millions of funded accounts no one can buy

@MaxForAI 如果 Kate 真去做 ChatGPT Web 基建，我更关心沙箱和恢复。能生成网站不稀奇，失败后还能接着跑才稀奇。

啊？？Cloudflare的Agents与Sandbox负责人Kate Reznykova宣布她加入了OpenAI，将参与ChatGPT的Web基础设施建设。 她此前负责Cloudflare Agents和Sandbox团队，参与开发Agents SDK、Sandbox SDK以及Project Think，核心工作就是让AI Agent拥有独立的代码执行环境、文件系统、持久状态和后台进程。 所以这次招聘可能不只是为了让ChatGPT网页打开得更快，而可能是为了ChatGPT的Web Agent布局？ 有意思的是今年OpenAI推出了Sites。 现在ChatGPT和Codex可以直接创建、保存、部署和托管网站、Dashboard、内部工具与Web应用，OpenAI同时提供Hosting、存储、数据库和权限控制。也就是说，从产生需求、写代码到部署上线，整个流程都可以留在ChatGPT里。 甚至连GitHub这一层，OpenAI也准备自己做。 之前据The Information和Reuters报道，OpenAI正在开发自己的代码托管平台。项目的直接原因之一，是GitHub多次故障影响了OpenAI内部工程协作，未来也可能向客户开放。不过这件事目前仍然只是媒体报道，OpenAI尚未正式发布。 随着ChatGPT逐渐从一个聊天产品变成能够长期运行任务的Agent，OpenAI也需要重新搭建它的Web工程底座。 Kate过去几年做的，恰好就是这件事。

@MaxForAI @vercel @nicoalbanese10 @OpenAI Nico 的经历确实很对口。Codex App 现在缺的不是再多一个模型按钮，是长任务怎么被看见、接管和恢复。

OpenAI又从 @vercel 挖来了一名Agent应用层核心成员。 Vercel AI SDK核心成员@nicoalbanese10 宣布加入 @OpenAI ，参与Codex App的开发。 他并不是传统意义上的模型研究员，也不是单纯做开发者关系的人。 他过去几年的工作，基本都围绕同一个问题展开：如何把底层模型能力，变成开发者每天真正能够使用的产品。 Nico本科就读于USC、香港科技大学和博科尼大学联合开设的World Bachelor in Business项目，毕业后没有直接进入科技公司，而是在伦敦的早期基金Ascension负责Pre-seed项目投资。 后来他开始自己写产品，并开发了Kirimase。 Kirimase是一套面向Next.js开发者的命令行工具，思路类似Ruby on Rails的脚手架。 开发者可以通过命令直接配置数据库、身份验证、Stripe支付、邮件服务，并生成完整的CRUD代码。这个项目获得了约2800个GitHub Star，后来被迁移到了Vercel官方GitHub组织下。 2024年前后，Nico加入Vercel，进入AI SDK核心团队。 过去两年里，他参与了AI SDK从3.3到6.0的多次关键版本升级，包括模型中间件、多步骤工具调用、结构化输出、类型安全的Chat、Agent循环、MCP、工具执行审批，以及Agent调试工具DevTools。 到2025年7月，Vercel AI SDK的周下载量已经超过200万。 这套SDK实际上定义了大量Web开发者今天构建AI应用的方式：如何接入不同模型，如何流式返回结果，如何展示工具调用，如何保存Agent状态，如何处理多步任务，以及如何调试模型在每一步到底做了什么。 最近，Nico还开发并开源了Open Agents。 这是一个运行在云端的Coding Agent系统，支持GitHub集成、持久化沙箱和可恢复的长时间工作流。 Nico称，这个Agent后来写下了他发布的每一行代码，甚至包括Open Agents自身。 项目开源后获得了超过5000个GitHub Star。 可以说Nico的经历和Codex App几乎完全对口。 Codex现在需要解决的核心问题，已经不只是模型能不能写出代码。 随着Agent能够同时运行几十分钟甚至数小时，产品还需要处理任务状态、文件系统、权限审批、GitHub协作、失败恢复、并行Agent，以及用户如何查看和接管Agent的工作。 这些恰好都是Nico过去两年在Vercel反复处理的问题。 他自己也表示，过去几个月一直全天使用Codex，自己发布的每一行代码和绝大部分其他工作，几乎都在Codex里完成。 现在，他从Codex的重度用户，直接变成了Codex App的建设者。 祝贺Nico！

@ChKashifAli Vertical AI doesn’t need to be the fastest to $100M. Tax workflow depth and trust are harder to copy than a generic agent demo.

Allow me to reintroduce myself “After the algo change” We started the first vertical AI startup in January 2023 before the term vertical AI and vibe coding existed. We are not the fastest to $100M ARR type of startup but we are changing the way taxes are done.

@SciTechera 3.9GB is impressive. I’d rather see sustained phone latency and battery drain than another retained-performance percentage.

Wow. Thats amazing! 27B AI MODEL NOW RUNS ON A PHONE 🤯 "PrismML just unveiled Bonsai 27B, the world's first 27B-class multimodal AI model designed to run locally on a smartphone." "It's built on Qwen3.6 27B, Bonsai 27B compresses a model that normally requires ~54 GB of memory down to just 3.9 GB, while reportedly retaining 90% of the original model's performance." "Instead of relying on cloud servers, it enables powerful on-device AI with offline reasoning, lower latency, better privacy and dramatically lower inference costs." "PrismML also released a 5.9 GB ternary version that retains 95% of the full-precision model's performance, making 27B-class AI practical on consumer laptops." This could mark a major milestone for edge AI, bringing powerful AI agents directly to smartphones and personal devices. Enjoy the acceleration!

@pixeluibygoogle Offline multimodal on Pixel 10 is useful if the handoff stays invisible. Which tasks remain fully local when Gemma 4 hits its limit?

Google announced Gemma 4, a lightweight multimodal AI model optimized for the Pixel 10’s Tensor G5 TPU, enabling offline chat, image recognition, audio transcription, and device control features across Pixel 10 devices. More details: developers.googleblog.com/unlocking-the-…

@CryptoTied 项目清单最有价值的不是“抄作业”，是看同一需求已经有多少人做了。找灵感前先找拥挤度。

🚨卧槽！中国独立开发者项目清单火了 这个仓库叫 chinese-independent-developer，目前已有 52k+ stars，是一个专门收集中国独立开发者作品的列表。 里面收录了大量实用、有趣的网站和 App 项目，很多都是个人或小团队做的（包含大量 AI 工具、效率工具、生产力产品等）。项目按时间排序，每条都有简洁介绍和链接，方便快速浏览。 核心价值： →发现国内独立开发者的真实作品 →看到别人在做什么（无论是赚钱项目还是玩票性质） →很多项目免费、开源或浏览器直用，实用性强 对想找灵感、了解国内 indie dev 生态、或者寻找好用的工具的人来说，这个列表非常值得收藏。

@milesdeutscher Cutting subscription costs is useful, but routing every task away from frontier models can create a review bill. Where did Hermes actually need the expensive model?

Hermes Agent killed the AI subscription model months ago. I don't get why more people aren't using it to cut their AI costs by 50%+. Stop paying monthly just to hit usage caps. Here's how to use Hermes to significantly cut your token spend: First, understand where your money is actually going. Most people overspend on AI in three places: 1. Defaulting to frontier models for every task that 2. Running everything through messaging gateways that bloat token counts 3. Starting from zero every session because there's no memory. Hermes fixes all three. 1. Stop defaulting to frontier models Hermes connects to 300+ models. I recommend setting DeepSeek V4 Flash as your default at $0.14 per million input tokens. Switch to Claude or GPT-5.6 only when the task actually needs frontier-level intelligence. Same output on 80% of tasks, yet 30x cheaper. 2. Use CLI over the gateway Running Hermes through Telegram or Discord sends 15,000 to 20,000 tokens of tool definitions per request. The CLI sends 6,000 to 8,000. Same result, but 2-3x cheaper per prompt. 3. Enable the learning loop Toggle on persistent memory and skill generation. Every complex task Hermes completes creates a reusable skill file. After 30 days, it completes similar tasks 40% faster, which means 40% fewer tokens are burned for the same work. With Hermes, costs drop the longer you run it - the opposite of every subscription you've ever had. 4. Set a VPS and forget it A Hetzner CX22 runs at $4/month. This allows your agent to run 24/7, hibernate when idle, and incur almost no overhead between tasks. The math: → Subscription model: $20/month, usage capped, memory wiped every session → Budget Hermes setup: $5 to $10/month, no caps, gets cheaper over time

@CNBCTV18Live Localized Claude pricing only works if the annual commitment still feels local. I’m curious whether monthly pricing changes too, not just the currency display.

#Anthropic starts localising #Claude pricing for #India, its biggest market after the US On Claude’s website in India, Anthropic is listing #ClaudePro at ₹2,000 (about $21) a month when billed annually, compared with $17 a month in the U.S.

@LuminaXspace A beta selector sighting is a useful signal, not a model review. I’d wait for pricing, limits, and a task-level eval before choosing Kimi K3.

🚨 Kimi K3 has reportedly started appearing on Moonshot AI’s official beta site Some users are now seeing Kimi K3 inside the beta model selector: • Kimi K3 reportedly visible on Moonshot’s beta website • Access appears limited to selected accounts • The model is not yet available on the main Kimi platform • Rumours suggest a 2–3T parameter architecture • Official benchmarks, pricing and context limits remain unknown This follows the leaked July 15 (Chinese Time) Kimi K3 launch promotion, suggesting the full release could begin within hours. Do you think Kimi K3 be the strongest Chinese frontier AI model?

@charles_lukes An audience helps only when it contains people with the problem. Serving software engineers is more specific than building a generic ‘builder audience.’

For all software engineers. Especially my guys in 🇳🇬. The new meta is building an audience and serving them with products they actually need. I know a lot of us just want to stay behind the keyboard, write code, and make money. I used to think the same way. But that era is changing. Today, almost anyone can build software with AI. Building better software is still a skill but the barrier to entry has dropped dramatically. That means writing code is no longer enough. If you want to stay ahead, you need distribution. You need people who know what you care about, trust your perspective, and are paying attention when you launch something. I’ve been building two AI products, and one thing has become very clear to me: For consumer apps especially, having an engaged audience is a massive advantage. Great products still matter, but great products with distribution win far more often. You can’t approach a software engineering career the same way you did three years ago. Come out of your shell. Share what you’re learning. Build in public. Teach. Help people. Find a niche you genuinely care about, and keep showing up. The code gets people to stay. The audience gets them through the door.

@Ulobex A polished Product Hunt deck in minutes is useful. I’d still spend the saved time tightening the positioning; templates can’t decide what the launch is actually about.

I wanted to see if Gamma's new templates could handle something beyond a typical slide deck. So I started with a rough AI startup idea and turned it into a launch-ready Product Hunt presentation in just a few minutes. The biggest surprise wasn't the speed. It was how polished the final deck looked with barely any editing. It genuinely felt like something I'd be comfortable presenting to investors or early users.