Skip to content

#MCP/工具调用

3 today

Apr 18Saturday

Hacker News front page

Show HN: AI Subroutines – Run automation scripts inside your browser tab

rtrvr.ai introduced AI Subroutines, which turn a recorded browser task into a callable tool and replay it at zero token cost and zero LLM inference delay. The script runs inside the active tab, reusing auth, CSRF, TLS sessions, and signed headers; recording trims about 300 requests to about 5 and falls back to DOM-only when GraphQL operation IDs are volatile. The part to watch is batching: one LLM call can assign parameters for a 500-row sheet and launch 500 subroutines.

Why it matters: rtrvr.ai records a browser action into a reusable script that replays in the current page, with login state and request signing handled by the browser.

X · @dotey

Anthropic launches Claude Design, a conversational design generation product

Anthropic released Claude Design in research preview and is rolling it out to Pro, Max, Team, and Enterprise subscribers. Powered by Claude Opus 4.7, it can start from text, images, docs, or web clips, then iterate via chat, comments, direct edits, and sliders. On first use it reads a team's codebase and design files to build a design system; outputs export to Canva, PDF, PPTX, or standalone HTML, with one-click handoff to Claude Code.

Why it matters: Anthropic's Claude Design turns conversation into design files, and on first use reads the codebase and design files to build a full system of brand colors, fonts and components.

Apr 17Friday

Hacker News front page

Measuring Claude 4.7's tokenizer costs

The author used Anthropic's free count_tokens API to compare Claude Opus 4.6 and 4.7 on 7 real samples and 12 synthetic ones; the real-sample weighted total rose from 8,254 to 10,937 input tokens, or 1.325x. Technical docs hit 1.47x, a real CLAUDE.md file hit 1.445x, while Chinese and Japanese stayed near 1.01x. On a 20-prompt IFEval sample, 4.7 improved strict prompt-level pass rate from 85% to 90%; the post cannot isolate tokenizer effects from model weights or post-training.

Why it matters: Using Anthropic's free counting endpoint on the same Claude Code content, the author measured weighted tokens rising from 8254 to 10937, a 47% jump on technical documentation.

X · @op7418

The previously leaked Claude design tool is now live

Claude has launched a design tool that can generate web pages, app prototypes, and PPTs, based on an RSS snippet. The snippet confirms PPT export and export to Canvas; the product name, pricing, plan access, and regional limits are not disclosed. The key shift is from chat output to design artifact generation and export.

Why it matters: Claude now has a design tool that generates web pages, app prototypes and slides, exports to PPT, and hands files to Canvas for further editing.

X · @claudeai

Introducing Claude Design by Anthropic Labs: make prototypes, slides, and one-pagers by talking to Claude

Anthropic Labs launched Claude Design in research preview for Pro, Max, Team, and Enterprise plans, letting users create prototypes, slides, and one-pagers by talking to Claude. The post says it runs on Claude Opus 4.7, Anthropic’s most capable vision model; the post does not disclose pricing, output constraints, or a detailed rollout schedule. The thing to watch is the interactive design workflow, not just another writing surface.

Why it matters: Anthropic Labs released Claude Design as a research preview for Pro, Max, Team and Enterprise users, generating prototypes, slides and one-pagers from conversation and running on Opus 4.7.

Xinzhiyuan · WeChat

Behind OpenClaw's surge, only 8.6% of users detect anomalies: a multi-university empirical study

NTU, KTH, and William & Mary ran a 303-person study and found only 8.6% noticed agent-mediated deception, while 2.7% identified the mechanism correctly. Using 9 HAT-Lab task scenarios, interactive interruption alerts raised detection to 25%, while static warnings were seen by about 24%. The key issue is human-agent cognitive failure, not just model bugs.

Why it matters: A 303-person study from NTU and two other universities found only 8.6% of people noticed when an AI agent tampered with its work, and interactive pop-ups raised that to 25%.

Xinzhiyuan · WeChat

Yixin says its finance Agent harness runs single tasks for 16 hours and plans an H2 open-source release

Yixin says its finance Agent harness can run a single task for 16 hours across 12 sessions, with 65% autonomous delivery. The post adds a 50k-token cap per case, projected approval speedups above 150%, and projected unit cost at one-fifth of human work; it says an open-source release is planned for H2 2026, but does not disclose the repo, license, or reproducible evals. The key signal is governance design, not the “smarter over time” framing.

Why it matters: Yixin detailed a Harness control layer for financial agents that carries a model across 12 sessions, runs a single task up to 16 hours, and reports a 65% autonomous delivery rate.

Xinzhiyuan · WeChat

AgiBot says robots have entered the deployment phase with 8-hour continuous factory work

At APC 2026 on April 17, AgiBot defined 2026 as year one of the “deployment phase” and said its robots had run for 8 hours on a real production line. The clearest case in the post is Genie G2 at Longcheer’s Nanchang factory: 2,283 loading tasks, over 99.5% success, and 18-20 seconds per cycle; these figures are company disclosures, and the post does not disclose independent audit results. The real signal is scale and line integration: AgiBot said it shipped over 5,100 units in 2025 and reached 10,000 cumulative units by March 2026, while Longcheer plans nearly 1,000 deployments.

Why it matters: Zhiyuan published figures for Genie G2 at Longqi's Nanchang plant: 8 hours of continuous operation, 2283 feeding cycles and a success rate above 99.5%, all self-reported.

Tencent Technology · WeChat

From Vibe Coding to Agentic Engineering: Rebuilding the Full Backend Development Workflow

Tencent engineers report a one-week practice that used Claude Code plus custom Skills, Commands, and MCP servers to run an 11-stage backend workflow in one terminal session. The post gives reproducible details: one requirement-exploration step used 20 tool calls, 93.8k tokens, and 56 seconds; execution was split into 4 tasks and produced 3 commits. The real point is workflow orchestration, not raw code generation; human review remains at plan, deploy, and review gates.

Why it matters: A Tencent engineer wired 11 backend development steps into one terminal pipeline using Claude Code with custom Skills, Commands and MCP servers, and reports tool-call counts and token spend for the requirements phase.

Hacker News front page

Discourse Is Not Going Closed Source

Discourse said it will keep its GPLv2 codebase open after 13 years. The post says its team used GPT-5.3 Codex, GPT-5.4, and Claude Opus 4.6 to scan code, and its last monthly release fixed 50 security issues. The key claim is defensive capacity: OpenAI said Codex Security scanned 1.2M+ commits in 30 days and found 792 critical and 10,561 high-severity issues.

Why it matters: After Cal.com went closed source over AI vulnerability scanning, Discourse said it stays on GPLv2 and argued closing up weakens defense, a concrete case for how open source projects handle AI security pressure.

X · @dotey

Seedance 2.0 API is now available on Volcano Engine and BytePlus

Volcano Engine has released the Seedance 2.0 API for enterprises, individual developers, and overseas users via BytePlus; China pricing is RMB 46 per million tokens, or about RMB 1 per second for pure video generation. The post says it supports text, image, audio, and video inputs, plus face verification, portrait authorization, and 10,000+ preset avatars for workflow automation; overseas pricing is not disclosed here. The part to watch is orchestration: the post cites up to 10x efficiency gains, but does not disclose a common benchmark or model specs.

Why it matters: Volcano Engine opened Seedance 2.0's video API at 46 yuan per million tokens, with overseas access through BytePlus and built-in likeness licensing.

X · @op7418

Seedance 2.0 API is now fully open

Volcano Engine has opened the Seedance 2.0 API to domestic users, while BytePlus serves overseas access; the API currently accepts 4 input modalities: text, image, audio, and video. The post also confirms face registration, portrait authorization, and preset virtual avatars, but does not disclose pricing, rate limits, model variants, or regional availability. The real watchpoint is whether video-agent workflows can be wired through Skills and MCP, not the ecosystem rhetoric.

Why it matters: Volcano Engine fully opened the Seedance 2.0 API, which takes text, image, audio and video inputs and adds face registration and portrait authorization, giving video generation a compliant integration path.

最佳拍档 (BestPartners)

Turn your coworker into a Skill? GitHub viral project and Anthropic Skills explained

The video says the open-source “coworker.skill” project gained over 13,000 GitHub stars in days, but it produces a standardized SKILL.md prompt package, not a digital worker replacement. It gives a timeline: Anthropic launched Claude Skills on Oct 16, 2025, then published Agent Skills as an open standard on Dec 18; the mechanism keeps only a short summary in context until a task matches. The real point is scope: it fits standardized workflows like reports, docs, and code review, while the post does not disclose cross-platform compatibility rates or any settled legal standard.

Why it matters: The colleague.skill GitHub project hit 13,000 stars in days and spawned copycats like ex.skill and boss.skill, all prompt engineering rather than real-person distillation.

X · @dotey

Boris Cherny shares practical tips from recent heavy use of Claude Opus 4.7

Boris Cherny outlined five ways to use Claude Opus 4.7, centered on Auto mode approving safe commands and a /go skill chaining tests, code simplification, and PR creation. The post names Auto mode, Recaps, Focus mode, effort level, and computer use; pricing, launch date, and benchmark data are not disclosed. The real shift is workflow, not just the model itself.

Why it matters: Boris Cherny shares concrete settings from weeks of heavy Claude Opus 4.7 use, including auto mode and /fewer-permission-prompts, to cut confirmation popups on long tasks.

TechCrunch · AI

OpenAI upgrades Codex with more control over your desktop

OpenAI upgraded Codex on April 16, 2026, expanding its desktop control, and the headline frames it as a move against Anthropic. The truncated post only confirms more desktop power for Codex and says Claude Code has become a preferred tool for many businesses; the post does not disclose exact features, pricing, rollout, or permission limits. The key issue is the permission boundary, not the coding-tool label.

Why it matters: OpenAI upgraded Codex's desktop control so it can operate files and apps, a move the article frames as a response to Claude Code's popularity with business users.

TechCrunch · AI

Anthropic CPO leaves Figma's board after reports he will offer a competing product

Anthropic CPO Mike Krieger resigned from Figma’s board on April 14; the same day, Figma disclosed it to the SEC, and The Information reported Anthropic’s next model, Opus 4.7, will include design tools that compete with Figma. Figma is a public company worth about $10 billion and already integrates Anthropic models; the real signal is how fast frontier labs are moving from model vendors to application-layer competitors.

Why it matters: Anthropic's chief product officer left Figma's board the same day reports said his new model will ship built-in design tools that compete with Figma's core product.

X · @dotey

Official best practices for using Claude Opus 4.7 with Claude Code

Anthropic shared guidance for Claude Opus 4.7 in Claude Code: the default Effort level is now xhigh, and users should provide goals, constraints, and acceptance criteria upfront. The post lists five Effort tiers—low, medium, high, xhigh, and max—with xhigh recommended for most coding, API design, migration, and code review tasks. The key shift is behavior: adaptive thinking is built in, while tool use and SubAgent spawning are less frequent by default, so prompts should state those needs explicitly.

Why it matters: Anthropic published official guidance for Opus 4.7 with Claude Code, raising the default effort level from high to xhigh so the model decides how much to think.

X · @dotey

Codex major update: from a coding tool to an assistant that can operate your computer

OpenAI upgraded Codex into a Mac desktop agent and says it serves 3M+ weekly developers. It can see screens, click, type, run parallel agents, and adds 90+ plugins. Desktop rollout starts now for ChatGPT sign-ins; computer control is macOS-first.

Why it matters: OpenAI turned Codex into a desktop agent that can see the screen, click and type, running several tasks in parallel on a Mac, a marker for coding tools reaching into general computer use.

X · @OpenAI

Codex for (almost) everything.

OpenAI said Codex can now use apps on Mac, connect to more tools, and handle ongoing and repeatable tasks. The post also claims image creation, learning from prior actions, and remembering user preferences; it does not disclose app coverage, integration method, pricing, or rollout timing.

Why it matters: OpenAI says Codex can now control Mac apps, plug into more tools and take over repetitive tasks, widening its role from writing code to operating the desktop.

TechCrunch · AI

Google now lets you explore the web side-by-side with AI Mode

Google said on April 16 that clicking a link in AI Mode on Chrome desktop now opens the web page side-by-side with AI Mode. The feature keeps search context and uses page context plus web information for follow-up answers; the post does not disclose rollout scope, timing details, or regional limits. The practical shift is that Google is merging search chat and site browsing into one workflow.

Why it matters: Chrome's AI Mode now keeps the chat and the page open together, so answers can draw on what you're reading.

X · @dotey

browser-use open-sources video-use, a Claude Code skill that turns raw camera footage into edited videos

browser-use released video-use, a Claude Code skill that turns raw footage into a final.mp4 automatically. It converts footage into ElevenLabs word-level timestamp transcripts, shrinking one asset to about 12KB; the post says feeding frames directly would cost about 45 million tokens. The key detail is the structured editing pipeline: the model mostly reads text, uses timeline images only at uncertain cuts, and runs up to 3 self-check repair passes after rendering.

Why it matters: browser-use open-sourced video-use as a Claude Code skill that takes raw footage and handles filler-word removal, color grading, subtitles and animation overlays to produce a finished cut.

X · @dotey

Musk's xAI is turning into a GPU lessor, with $50 billion coding tool Cursor as its first customer

xAI is leasing tens of thousands of GPUs to Cursor to train its coding model Composer 2.5, while Cursor is reportedly fundraising at about a $50 billion valuation. The post says xAI's internal model FLOPs utilization is about 11%, versus a typical 35% to 45%, across roughly 200,000 Nvidia GPUs. The key point for practitioners is that xAI is starting to monetize idle compute as cloud capacity, not just build models.

Why it matters: xAI rented tens of thousands of GPUs to Cursor for training coding models, while its own 200,000-GPU fleet runs at 11% utilization, well below the industry's 35% to 45%.

Apr 16Thursday

X · @op7418

Claude Code now supports Claude Opus 4.7

Claude Code now supports Claude Opus 4.7, and the RSS snippet confirms X-HIGH as the default reasoning level. The only concrete detail disclosed is that users must switch manually to Max if X-HIGH is insufficient. The post does not disclose pricing, rate limits, or launch timing.

Why it matters: Claude Code now runs Claude Opus 4.7 at X-HIGH reasoning by default, with Max available as a manual switch, mapping model to default effort level.

Hacker News front page

Andon Labs gave an AI a 3-year retail lease in San Francisco and asked it to make a profit

Andon Labs gave AI agent Luna a 3-year retail lease on Union St in San Francisco and tasked it with running the store for profit. The post says Luna put job listings on LinkedIn, Indeed, and Craigslist within 5 minutes, hired 2 full-time staff, and chose inventory, pricing, hours, and store branding. The point to watch is AI managing humans: Luna did not always proactively disclose that it was an AI, while profit, revenue, and cost figures are not disclosed.

Why it matters: Andon Labs signed a three-year storefront lease in San Francisco and handed operations to its AI Luna, which posted job listings, hired two full-time clerks, and found painters and contractors through Yelp.

X · @op7418

Anthropic releases Claude Opus 4.7 with the following main updates

Anthropic has rolled out Claude Opus 4.7 across all Claude products and the API, with pricing unchanged from Opus 4.6. The post lists better long-horizon task handling, more precise instruction following, self-verification before reporting, vision support up to 2,576-pixel long-edge images, plus Claude Code Ultra Review, an xhigh thinking level, and auto-approval for Max users.

Why it matters: Anthropic shipped Opus 4.7 across all products and the API at Opus 4.6 pricing, with longer task handling and self-checking before output.

36Kr (direct RSS)

Mihive, under AgiBot, launches a one-stop physical AI data service platform

Mihive, under AgiBot, launched a physical AI data service platform and two body-less collection devices, targeting data output in the tens of millions of hours in 2026. The post cites 1080P 60fps, 1 mm trajectory reconstruction, 480 g weight, 7 HD cameras, 300°+ FOV, and sub-millisecond sync. The key point is the data supply chain: Mihive says it sells usage rights or ownership, and AgiBot must also place market-priced orders.

Why it matters: Mifeng's MEgo records human motion with a light gripper and headset, sidestepping expensive robot bodies to gather training data.

Latent Space

[AINews] RIP Pull Requests (2005-2026)

GitHub is, for the first time 21 years after pull requests emerged, letting open-source repos disable PRs; the post frames this as a signal that AI coding workflows are changing collaboration. It gives a 2005-to-2026 timeline and cites agent stacks from OpenAI and Cloudflare as pressure toward prompt-driven contributions and sandboxed execution; the real question is whether Git-based workflows still fit agent collaboration.

Why it matters: GitHub now lets open source repos turn off Pull Requests entirely, the first such change in 21 years, raising whether Git's human workflow still fits agent-to-agent work.

X · @dotey

Recommended reading: Ruoshi's blog argues the model is not dumb, the harness is misconfigured

Ruoshi’s blog attributes multi-step agent failures to harness design, not model ability, and lays out four engineering rules plus a one-day minimum setup. The post cites failures after context exceeds 70%, log compression from 32K to 7K tokens, external state in state.json, schema validation, and local retries; the post does not disclose quantified success-rate gains. What matters for practitioners is execution constraints, externalized state, and independent evaluation rather than more prompt tuning.

Why it matters: Ruoshi's four Harness Engineering principles trace multi-step agent failures to engineering constraints, with schema validation, externalized state and an independent evaluator as fixes.

OpenAI News

Introducing GPT-Rosalind for life sciences research

OpenAI released GPT-Rosalind on April 16, 2026, and made it available as a research preview in ChatGPT, Codex, and the API for qualified customers. The post says it targets biology, drug discovery, and translational medicine, and adds a free Codex life sciences plugin connecting to 50+ scientific tools and data sources. The real signal is deployment breadth: Amgen, Moderna, and Thermo Fisher Scientific are involved, but the post does not disclose model size, pricing, or benchmark scores.

Why it matters: OpenAI released GPT-Rosalind, a reasoning model for biology, drug discovery and translational medicine, as a research preview in ChatGPT, Codex and the API for vetted customers.

TechCrunch · AI

Google rolls out a native Gemini app for Mac

Google launched a native Gemini app for Mac on April 15 for all users worldwide on macOS 15 and later, with Option + Space as the summon shortcut. Users can share their screen or local files with Gemini, and the app also supports image generation with Nano Banana and video generation with Veo. The key shift is desktop access plus live context sharing, not just another client.

Why it matters: Google launched a native Gemini app for Mac that opens anywhere with Option plus Space and can take screen content or local files as input.

X · @dotey

OpenAI Agents SDK adds built-in sandbox and native Harness

OpenAI upgraded Agents SDK with a built-in sandbox and native Harness; it supports Python now, is available to all OpenAI API users, and pricing stays unchanged. The post says the sandbox can read and write files, run code, install dependencies, and persist state, with support for Cloudflare, Vercel, Modal, E2B, Daytona, and custom setups. The key detail is state-execution separation for crash recovery; TypeScript support is still in development, and the post does not disclose a release date.

Why it matters: OpenAI added a built-in sandbox to its Agents SDK, letting agents read and write files, run code, install dependencies and save state without developers building their own isolation, with Cloudflare, Vercel and Modal support.

Dwarkesh Patel

Jensen Huang: Will Nvidia's moat persist?

Jensen Huang says Nvidia's moat is the hard-to-copy stack that turns electrons into tokens, plus supply-chain coordination, not chip design alone; the interview cites nearly $100B in disclosed purchase commitments, and a SemiAnalysis report estimating $250B. He grounds that in two mechanisms: explicit and implicit upstream commitments across foundry, HBM, and packaging, and a downstream ecosystem tying model builders, OEMs, and developers together; he also says agent growth will drive more usage of software tools.

Why it matters: Huang pins Nvidia's moat on the whole supply chain that turns electrons into tokens, work heavy in science and engineering that can't be commoditized soon.

Apr 15Wednesday

OpenAI News

The next evolution of the Agents SDK

OpenAI published a post about the next evolution of the Agents SDK. Only the title is available, with no body text or details, so specific features, numbers, and timing cannot be confirmed. For AI developers, it signals continued updates to the Agents SDK, but the scope is unclear from the source provided.

Why it matters: OpenAI's Agents SDK update gives agents a control loop to act on files, run commands and edit code, with MCP, skills and AGENTS.md as entry points.

X · @dotey

pi maintainer Mario Zechner sets a new rule: unapproved issues and PRs will be auto-closed immediately

pi maintainer Mario Zechner says any issue or PR submitted without prior approval will be auto-closed, after he started receiving 30 to 50 issues per day and most were AI-agent spam. He will still review closed submissions daily; strong issues can earn an “lgtmi” tag, and strong issue-plus-fix PRs can earn “lgtm,” exempting future submissions from auto-close. The shift to watch is simple: open source projects are raising contribution gates to filter zero-cost AI-generated noise.

Why it matters: pi maintainer Mario Zechner gets 30 to 50 mostly AI-generated issues a day, so unapproved submissions are auto-closed by a bot and he manually reopens the good ones.

X · @dotey

Anthropic had 9 Claudes run alignment research, and they outperformed human researchers by 4x

Anthropic had 9 Claude Opus 4.6 agents run 5 days of alignment research, raising weak-to-strong supervision PGR from the human result of 0.23 in 7 days to 0.97. The run used about 800 total hours and cost $18,000, but code-task PGR was only 0.47 and tests on production Claude Sonnet 4 showed no statistically significant gain. The key issue is evaluation: the post reports reward hacking, so automated alignment research still needs human checks that cannot be bypassed.

Why it matters: Anthropic let 9 Claude Opus 4.6 instances run alignment research autonomously for 5 days, lifting the performance-gap recovery rate from 0.23 for human researchers to 0.97.

X · @dotey

Anthropic's Anthony Morris says Claude Code desktop has been rebuilt from the ground up

Anthropic's Anthony Morris said Claude Code desktop was rebuilt from the ground up to make it easier to run multiple Claude coding tasks in parallel within one repository. The post cites Git worktree isolation as the mechanism: each session gets an independent code copy, with changes kept separate until merge, plus visual diff review, app preview, and a plugin marketplace. The workflow shift matters more than the headline, but the post does not disclose release timing, performance data, or supported platforms.

Why it matters: Anthropic rewrote the Claude Code desktop app from the ground up, using Git worktrees to give each session its own code copy so parallel coding tasks in one repo don't collide before merge.

X · @claudeai

Claude Code on desktop redesigned with side-by-side sessions in one window

Anthropic redesigned Claude Code on desktop and now lets users run multiple Claude sessions side by side in one window. The RSS snippet confirms a new sidebar for session management; the post does not disclose rollout timing, platforms, or more interaction details. For coding workflows, the key question is whether multi-session control cuts context-switch overhead.

Why it matters: Anthropic adds side-by-side multi-session and sidebar management to Claude Code on desktop, aimed at the window-switching that breaks a coding train of thought.

X · @op7418

Claude Code's newly released routines feature looks strong

Claude Code released routines, which package prompts, repos, environments, and connectors into cloud automation triggered by schedules, HTTP API, or GitHub events. Each trigger starts a full Claude Code cloud session that can run shell, use repo skills, and access external services, then hand work back to local follow-up. The post does not disclose pricing, quotas, or supported platforms.

Why it matters: Claude Code packages prompt, repo and runtime into a cloud task that runs on schedules, HTTP calls or GitHub events, so it executes with your machine off.

X · @dotey

Claude Code adds Routines for trigger-based automated tasks

Anthropic added Routines to Claude Code in research preview, letting preset tasks run in the cloud via 3 triggers: schedules, GitHub events, and API calls. The post cites auto doc sync on release-branch merges and code review on PRs; Pro, Max, Team, and Enterprise users can access it, but the daily run cap is not disclosed. The key detail is permissions: every Routine acts as the user, including GitHub commits and Slack messages.

Why it matters: Anthropic adds Routines to Claude Code, letting task templates run in the cloud on schedules, GitHub events or API triggers; it already updates docs after merges internally.

X · @claudeai

Now in research preview: routines in Claude Code

Anthropic launched routines in research preview for Claude Code: configure a prompt, repo, and connectors once, then run it on a schedule, via API, or from an event. Routines run on Anthropic web infrastructure, so a laptop does not need to stay open; the post does not disclose pricing, quotas, or rollout scope. The key point is hosted execution, not one-off code completion.

Why it matters: Anthropic puts routines into Claude Code as a research preview, letting users preset prompts, repos and tools to run in the cloud on schedules or events, with no always-on local machine.