Claude plugins, safer agents, and local voice AI
Anthropic has open-sourced 11 editable Claude plugins for work in sales, finance, support, and more. We also look at server-enforced agent approvals, faster WhatsApp Business setup, local voice tools, and several launches that still call for a cautious pilot.
The Big Story
GitHub Trending
knowledge-work-plugins: plugins for knowledge workers to use in Claude Cowork
Anthropic open-sourced 11 role-based plugins for Claude Cowork, with compatibility for Claude Code. The collection covers productivity, sales, customer support, product management, marketing, legal, finance, data, enterprise search, bio research, and plugin management, with connectors for tools such as Microsoft 365, Slack, HubSpot, Jira, Notion, BigQuery, and Canva. Each plugin packages skills, slash commands, sub-agents, and MCP connector settings as editable Markdown and JSON files, so organizations can add their terminology and workflows without a build or infrastructure layer.
Businesses already using Claude can start with one narrow, repeatable workflow and customize the relevant plugin, but should review connector permissions and test outputs before allowing it to act on customer, legal, or financial data.
For Everyone
Google Gemini Blog
Create your best tracks yet with Lyria 3.5 in Gemini.
Google released Lyria 3.5 globally in the Gemini web and mobile apps, with access also available through the Gemini API, Google AI Studio, Google Flow Music, and Google Vids. The music-generation model supports vocal and instrumental tracks, genre descriptions, starter templates, and a choice of short or longer outputs. Google positions it for uses such as video backing tracks, brand jingles, birthday songs, and ringtones.
Businesses that regularly need simple original music can test Lyria 3.5 for low-stakes marketing assets, but should confirm API pricing, commercial-use terms, and rights protections before publishing generated tracks.
Google Gemini Blog
4 ways to tackle household chores with Gemini
Google highlighted four household uses for Gemini: troubleshooting repairs with Gemini Live’s camera, creating meal plans and Instacart shopping carts, extracting spending data from receipt photos, and comparing major purchases. Google’s AI & Economy ATLAS v1.0 study analyzed 15 million de-identified interactions and found that more than 86% of conversational AI use happens outside formal work. Researchers estimated that saving each household 30 minutes per week could represent roughly $100 billion in annual U.S. economic value.
The same features can handle small-business chores such as diagnosing equipment, organizing receipts, planning supplies, and comparing purchases, but sensitive financial or customer information should stay out of uploads unless the account’s data controls meet the business’s requirements.
For Builders
AI TLDR
Claude Managed Agents add auto mode: the server checks every tool call
Anthropic added an `auto` permission policy to Claude Managed Agents on September 10, 2026. For every agent or MCP tool call, Anthropic's server decides whether to run it, deny it, or pause for human approval, and records that decision in tool-use events. Its ant CLI can also connect to a live agent session so an operator can watch it, send messages, and approve or deny waiting calls from a terminal or local web viewer.
If you're building a Claude-powered agent that can touch business systems, test whether these approval controls cover the actions you'd otherwise need to supervise yourself.
GitHub Trending
voicebox: The open-source AI voice studio
Voicebox is a free, open-source voice studio that runs voice cloning, speech generation, dictation, and stored audio locally rather than sending them to a cloud service. It supports seven text-to-speech engines, more than 50 preset voices, 23 languages, Whisper transcription, scripts up to 50,000 characters, audio effects, and a multi-track editor. REST and MCP interfaces let software and AI agents generate speech, while desktop builds are available for macOS and Windows, Docker is supported, and Linux currently requires a source build.
Businesses handling sensitive recordings or wanting predictable usage costs can pilot Voicebox as a local alternative to subscription voice services, but should test voice quality, hardware requirements, consent procedures for cloning, and platform-specific installation before broader use.
GitHub Trending
colibri: Run frontier MoE models on hardware you already own
Colibrì is a dependency-free, pure-C inference engine that runs mixture-of-experts models from 7 billion to 2.8 trillion parameters by treating VRAM, RAM, and NVMe storage as one memory hierarchy. For a 744-billion-parameter model, it keeps about 9.9 GB of dense components in RAM and streams roughly 370 GB of routed experts from disk, using caches and prefetching to move frequently needed weights into faster tiers. The project supports CPU, CUDA, Metal, and local Mac clusters, but explicitly offers no speed guarantee; one six-RTX-5090 example reached four tokens per second, while lower-residency systems trade speed for accessibility without silently changing model semantics.
This is an experimental option for technically capable businesses that need private local model access and can tolerate slow or hardware-dependent performance; most small firms should compare its total equipment and operating cost with a managed API before committing.
For Business
TechCrunch AI
Meta now lets AI agents handle the boring parts of WhatsApp Business setup
Meta released WhatsApp Business Tools MCP, which lets AI coding agents such as Claude, Cursor, Codex, and ChatGPT set up and manage WhatsApp Business messaging through conversational instructions. An agent can create an account, add and verify a phone number, register Cloud API access, check terms and business verification, create message templates, test messages and webhooks, and flag configuration failures. Meta’s separate Social Technologies MCP can help search documentation and troubleshoot API errors.
Businesses planning a WhatsApp customer-service channel can use this to reduce developer setup work, but should require human approval for phone verification, payment settings, message templates, and any action involving customer data.
TechCrunch AI
Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
Salesforce introduced Koa, its first reasoning model, built with Nvidia on the open-weight Nemotron model and post-trained for sales, marketing, and customer-support tasks. Koa will join Agentforce as an alternative to Claude and ChatGPT for multi-step work, with Salesforce claiming lower token use and enforcement of customers’ existing data and security requirements. Its training used synthetic sales and service scenarios rather than Salesforce customer data, while Salesforce will continue offering outside models through products such as Claudeforce.
Salesforce customers should compare Koa against their current Agentforce model on task accuracy, token cost, and data-governance requirements once pricing and availability are clear, rather than assuming a specialized model is automatically the better choice.
Worth Watching
AI News
Pony.ai unveils autonomous electric truck for logistics fleets
Pony.ai and GAC Commercial Vehicle unveiled a Level 4 autonomous electric heavy truck, with mass production planned later this year for long-haul routes, dedicated freight corridors, and ports. Its sensor kit uses nine lidars, three radars, and 13 cameras, with redundant steering, braking, power, computing, sensing, and communications. Pony.ai says hardware cost is 70% lower than its prior generation and projects 30% lower operating cost per ton-kilometer and 10% lower energy use, though these are company estimates rather than fleet results.
This isn’t ready for most small fleets, particularly outside its initial markets. Logistics operators should track real-world safety approvals, route coverage, maintenance support, and verified total cost before considering a pilot.
TechCrunch AI
Mecka AI nears $500M valuation in Sequoia-led deal amid rush for robot training data
Robotics data startup Mecka AI is nearing a Sequoia Capital-led funding round at a valuation of about $500 million, three months after raising $60 million. Founded in 2024, the company pays people to record everyday physical tasks with body sensors and smartphones, producing data used to train robots. Mecka projected a $100 million annual revenue run rate by the end of 2026, although the new financing terms remain subject to change.
This is an early infrastructure bet with no immediate small-business product to adopt, but it signals growing investment in robots trained for routine physical work.