Welcome to the daily AI news brief for vibe coders. It is Monday, July 13, 2026, and the agent tooling layer moved more than the model layer over the weekend. Anthropic pushed MCP tunnels into research preview and made Claude Sonnet 5 the default in Claude Code, Google opened Managed Agents in the Gemini API, and Genkit finally shipped Python support. If you build on Shopify, Claude, GPT, Gemini, or Firebase, the changes worth your Monday are all in the plumbing, not the leaderboards.
TL;DR
- MCP tunnels are in research preview, and the 60-second MCP timeout bug is fixed.
- Claude Sonnet 5 is now the Claude Code default with a 1M-token context window.
- Gemini Managed Agents run stateful agents in Google-hosted Linux sandboxes.
- Genkit 1.22 adds early Python support and expands Go.
- GPT-5.6 Sol, Terra, and Luna are fully public across ChatGPT, Codex, and the API.
- Apple filed a trade secret suit against OpenAI on July 11.
Anthropic ships MCP tunnels and fixes the 60-second timeout
What shipped. Anthropic moved MCP tunnel management from the Admin API to /v1/tunnels on the Claude API, gated behind the mcp-tunnels-2026-06-22 beta header and a workspace:manage_tunnels scope. In the same window, Anthropic fixed a bug where MCP servers configured through --mcp-config or .mcp.json ignored a per-server request_timeout_ms, silently falling back to the 60-second default in fresh sessions.
Why it matters for vibe coders. If a Shopify bulk operation or long Firestore migration ever died at exactly one minute inside an agent run, that was the bug. Per-server timeouts now actually apply.
What to do today. Audit every .mcp.json in your repos. Set an explicit request_timeout_ms on any server that talks to a slow API, then re-run the agent task that used to fail. Claude Code changelog
Claude Sonnet 5 becomes the Claude Code default
What shipped. Claude Sonnet 5 is now the default model in Claude Code, with a native 1M-token context window and promotional pricing of $2 in and $10 out per million tokens through August 31. You need version 2.1.197 or later to get it.
Why it matters for vibe coders. A million tokens of native context changes how you scope agent tasks. A whole Shopify theme — sections, snippets, schema, Liquid — fits in one window. You stop writing retrieval glue and start handing the agent the actual repository.
What to do today. Upgrade, confirm your model, and retry one refactor you previously had to split across sessions. The promotional pricing has an end date, so run your cost math against the post-promo rate before you build a product on it. Anthropic release notes
Google opens Managed Agents in the Gemini API
What shipped. Managed Agents launched in public preview in the Gemini API. They let developers build and deploy autonomous, stateful agents that execute inside secure, isolated Google-hosted Linux sandboxes. Gemini 3.5 Flash also reached general availability and now backs gemini-flash-latest.
Why it matters for vibe coders. This is Google’s answer to hosted agent runtimes. The interesting part is not the model — it is that Google now runs the sandbox, the state, and the lifecycle. For a solo builder, that removes the container work that usually kills an agent side project in week two.
What to do today. Stand up one Managed Agent with a narrow job: read a product feed, flag missing alt text, write results to a file. Compare it against the same agent on the Claude Agent SDK. Gemini API changelog
Genkit 1.22 adds Python and expands Go
What shipped. Genkit was upgraded to 1.22 across its agent packages, with early Python support and expanded Go support so you can build agentic flows outside JavaScript. Separately, note that the Firebase extension for Gemini CLI stopped working on June 18 as Gemini CLI moved to Antigravity CLI.
Why it matters for vibe coders. Firebase-backed agents were effectively a TypeScript-only story. If your data pipeline is Python — which, for most people doing anything with embeddings or scraping, it is — you no longer need a Node shim in the middle.
What to do today. Port one Genkit flow to Python and check your Firestore triggers still fire as expected. Do not assume parity with the JS SDK yet. Genkit releases
Also worth noting
- OpenAI’s GPT-5.6 family is fully public: Sol for frontier work, Terra for balanced cost, Luna for high-volume jobs, live across ChatGPT, Codex, and the API. TechCrunch
- GPT-Live-1 and GPT-Live-1 mini are full-duplex voice models that listen and respond continuously instead of waiting for a turn boundary. Axios
- Gemini 3.5 Pro is reported to reach general availability on July 17. Reported, not confirmed by Google — treat the date as unverified.
- Apple filed a trade secret suit against OpenAI in Northern California federal court on July 11, centered on former Apple staff now working at OpenAI. Roundup
Build of the day
Fix your MCP timeouts, then prove it. Open the .mcp.json in your Shopify project, set request_timeout_ms to 300000 on your Shopify Admin MCP server, and give Claude Code one long-running task you previously abandoned: pull every product in the catalog, check each one for a missing image alt attribute and a missing product type, and write a CSV of the offenders. Before the fix, that job died at sixty seconds on any catalog past a few hundred SKUs. Time it. If it finishes, you have a confirmed fix and a data quality report you needed anyway. Two hours, including coffee.
Sources
- https://code.claude.com/docs/en/changelog
- https://releasebot.io/updates/anthropic/claude-code
- https://ai.google.dev/gemini-api/docs/changelog
- https://github.com/firebase/genkit/releases
- https://techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/
- https://www.axios.com/2026/07/09/ai-openai-gpt-release
- https://unrot.co/blogs/today-top-10-ai-news-july-12-2026
- https://medium.com/@davidakpovi/ai-news-week-of-july-6-to-july-12-2026-f81a26c49c55
FAQ
What are MCP tunnels and do I need them?
MCP tunnels are in research preview. Management moved to /v1/tunnels on the Claude API and requires the mcp-tunnels-2026-06-22 beta header plus a workspace:manage_tunnels scope. If you run MCP servers that a hosted Claude deployment needs to reach, tunnels are the sanctioned path. If your servers run locally alongside Claude Code, you do not need them yet. See the Claude Code changelog.
Why do my MCP tool calls time out at 60 seconds?
That was a real bug, and it is fixed. MCP servers configured via --mcp-config or .mcp.json were ignoring a per-server request_timeout_ms and silently reverting to the 60-second default in fresh sessions. Update Claude Code, then set the timeout explicitly on any server that calls a slow API. Long Shopify bulk operations were the classic casualty. Source: Claude Code changelog.
Is Claude Sonnet 5 worth switching to for Shopify theme work?
It is already the default in Claude Code, so the switch is made for you on version 2.1.197 and later. The native 1M-token context window is the practical win: an entire theme, including sections, snippets, and schema, fits in one window without retrieval scaffolding. Promotional pricing of $2 and $10 per million tokens runs through August 31. Details: Anthropic release notes.
What are Gemini Managed Agents used for?
Managed Agents, now in public preview in the Gemini API, run autonomous, stateful agents inside secure, Google-hosted Linux sandbox environments. The value for a solo builder is operational: Google runs the sandbox and the session state, so you skip the container and lifecycle work. Good fit for scheduled catalog jobs and background data cleanup. See the Gemini API changelog.
Can I build Firebase agents in Python now?
Yes, with a caveat. Genkit 1.22 introduced early Python support and expanded Go support across its agent packages, so agentic flows are no longer JavaScript-only. “Early” is the operative word — do not assume feature parity with the JS SDK, and test your Firestore triggers after porting. Also note the Firebase extension for Gemini CLI stopped working on June 18. Source: Genkit releases.
Which GPT-5.6 variant should a small store pick?
OpenAI ships three: Sol for frontier capability, Terra for a balance of intelligence and cost, and Luna for efficient, high-volume workloads. For a small store, Luna handles the bulk jobs — product descriptions, alt text, tagging — and Terra handles anything that needs judgment. Reserve Sol for hard reasoning you cannot batch. Details: TechCrunch.
About the author
Robert McCullock builds AI-assisted commerce systems at Design Delight Studio, working across Shopify Liquid, React, TypeScript, Python, Firebase, and multi-agent architectures on the Claude Agent SDK. See his work at the Design Delight Studio portfolio.