Announced Nov 6 for Pro, Max, and Ultra, Claude works out the dependencies, network access, and API keys a repo needs in a cloud session, then saves that environment for next time. Keys are stored as network secrets that Anthropic's proxy attaches after requests leave the VM, so Claude never sees them.
Google Cloud's Data Agent Kit is now generally available as a free set of MCP tools and open source skills that let Claude Code, Codex, Antigravity, and VS Code based editors inspect schemas, run queries, and read job logs across 15+ Google data services. GA adds BigQuery Graph, Bigtable, and Managed Service for Apache Spark, and the agent runs with your own IAM permissions…
Available now through Antigravity's Science plugin or npx skills add google deepmind/science skills/ , the skills let the model orchestrate while RDKit and the ChEMBL and PubChem connectors do the deterministic chemistry. DeepMind's demo has an agent sift thousands of compounds and propose macrocycle designs in about an hour, and it stresses every design still needs lab validation…
As of Nov 4, CORTEX AGENT RUN and CODING AGENT tools accept background: true and return a run handle at once, which you poll with the auto added ASYNC AGENT RUN STATUS , AGENT RUN RESULT , and ASYNC AGENT RUN CANCEL tools. Snowflake's REST docs say a background run can continue for up to 6 hours, versus a 15 minute default timeout for synchronous requests…
Published Nov 2, Glow Labs found that in 98% of sessions where an agent set up a deprecated MCP server (e.g. @modelcontextprotocol/server github instead of GitHub's official one), the name came from training data, across Claude Code, Cursor, and Codex CLI. Four high severity bugs include a prompt injection to RCE in the archived Puppeteer server, which maintainers won't fix; scan your MCP configs and pin approved server packages…
The Apache 2.0 desktop app for macOS, Windows, and Linux exposes a USB or Wi Fi connected phone to Claude Code, Codex, Cursor, or any MCP client via pinkyctl mcp , and its gateway, not the prompt, holds payments and other flagged taps until you click Allow. It's pre alpha, with Claude Code the only agent fully tested on real phones, and iOS 17.4 to 26 needs a WebDriverAgent helper built from Xcode…
Spotted something we missed? Start a thread.
