WebMCP puts agents inside websites
Your AI stack is getting more hands-on. Today’s updates move agents into websites, local machines, your existing work context, and the apps your team already uses.
In today’s edition
- WebMCP lets websites expose page-level tools that ChatGPT and Codex can use to complete tasks.
- Perplexity launched a fully local agent runtime for NVIDIA DGX Spark users.
- Vercel Connect is generally available with secure, scoped access to more than 100 services.
- Claude now shares one editable memory across chat and Claude Cowork.
- ElevenLabs released Composer, a song editor that lets creators revise tracks section by section.
WebMCP gives websites an agent interface
WebMCP is an experimental open standard that lets web apps expose tools agents can use directly on the page.
- ChatGPT’s desktop app browser and ChatGPT Sites are adding WebMCP support.
- On compatible sites, ChatGPT or Codex can use the site to complete a task.
- Builders can start from scratch or add WebMCP to an existing site.
- The WebMCP Challenge offers $35,000 in cash prizes, with submissions due September 3 at 5pm PT.
What it means Add task-specific tools to your web product now, so ChatGPT and Codex can act through your interface instead of relying only on clicks and screen reading.
Source: OpenAIDevs
Perplexity ships a fully local agent runtime
Perplexity launched Portable Computer on NVIDIA DGX Spark, putting its entire Computer runtime on local hardware with no cloud dependency.
- The local setup includes the orchestrator LLM, subagent LLM, and agent harness.
- It runs a locally installed, post-trained PPLX 27B model, with Qwen 3.8 27B also available.
- When frontier reasoning is needed, the system asks for approval before routing text guidance to a cloud model.
- Portable Computer is available to Perplexity Pro and Max subscribers with an NVIDIA DGX Spark.
What it means Test local-first agent workflows for private documents and controlled costs, while keeping a user-approved cloud escalation path for harder tasks.
Source: perplexity_ai
Vercel Connect opens secure app access
Vercel Connect is now generally available, giving apps and agents secure access to Slack, Linear, GitHub, and more than 100 other services.
- It uses short-lived, scoped access tokens.
- Builders get token and trigger observability.
- RBAC and audit trails are included.
- v0 can use Vercel Connect directly with services including Slack, GitHub, Notion, and Salesforce.
What it means Use scoped tokens and audit trails instead of building connector security from scratch when your agent needs to work across customers’ existing software.
Source: vercel
Claude memory now follows work across modes
Claude now uses one memory across chat and Claude Cowork, so a Cowork task can begin with context from your prior conversations.
- Memory can carry over project details, manager preferences, and past client context.
- You can read, edit, or delete each saved topic in Settings.
- You can explicitly save details by saying “remember this.”
- Sensitive topics such as health or religious beliefs stay out of memory unless you enable them in Settings.
What it means Set up Claude’s memory around recurring clients, projects, and preferences to reduce briefing time, then review saved topics before handing it ongoing work.
Source: claudeai
ElevenLabs makes AI songs editable by section
ElevenLabs introduced Composer, a section-by-section song editor for starting from lyrics, a track, a blank page, or a prompt.
- Upload an existing track and open it as editable verse, hook, and bridge blocks.
- Rewrite one verse, change a chorus key, or slow the bridge without changing the rest of the track.
- You can adjust the duration, tempo, and key of any section.
- Composer is powered by Music v2 and available in ElevenMusic.
What it means Use section-level generation to produce client-ready music faster, because you can revise the weak moment of a track without regenerating the whole thing.
Source: ElevenLabs