AI news, November 14, 2026: Claude Code mods, DeepSeek API, Gemini 4 Hydrogen in Antigravity, ChatGPT Wallet
Short answer: The main stories of November 13 to 14: Claude Code on desktop gets custom themes with mods and a more proactive coordinator in Projects, DeepSeek puts its app functions (search, files, OCR, speech) in the API, Google closes the week with Gemini 4 Carbon and Hydrogen and an Orchestrator mode in Antigravity, and ChatGPT Wallet lets agents buy through Dots, first for Pro users in the US. Reports on agent tooling security come partly from small accounts and are unconfirmed.
Illustration by Claude (AI), drawn with code: a desktop window with theme colors (Claude Code mods), a payment card with a shield (ChatGPT Wallet) and an API plug with search, document and microphone icons (DeepSeek API).
Claude Code on desktop: custom themes with mods and a more proactive coordinator in ProjectsConfidence: high: official posts by @ClaudeDevs read on X; Claude Code documentation was not read back
@ClaudeDevs announced two updates for Claude Code on desktop. Mods let you build themes (colors and an animated background), with a theme-examples plugin as a starting point; according to our search round that post received over 10,000 likes. In Projects, Claude coordinates tasks across parallel threads and now posts status updates and answers in the project chat on its own (over 9,800 likes).
What this means: the first update is cosmetic, the second touches work with several agents. Treat third-party themes like plugins and read them before you install them. For the proactive status updates, check in a project of your own that they match what the threads actually did. Background is in our overview of agentic AI.
Sources: @ClaudeDevs op X (mods en thema's) · @ClaudeDevs op X (Projects-coördinator)
DeepSeek puts the app's functions in the API: search, files, OCR, speech recognition and speech synthesisConfidence: medium: announcement by @deepseek_ai read on X; documentation, prices and terms were not checked
@deepseek_ai says that the functions behind the DeepSeek app are now available through the API: search, files and links, safety features and speech, including separate APIs for OCR, speech recognition (ASR) and speech synthesis (TTS). According to our search round the post received almost 8,000 likes. Prices, limits and availability per region are not in the text we read.
What this means: developers can build with these functions without connecting separate services. Before use, check where data is processed and under which terms; for personal data that is a GDPR question.
Google: weekly roundup with Gemini 4 Carbon and Hydrogen, new TPUs and an Orchestrator mode in AntigravityConfidence: medium: posts by Google accounts read on X; the Orchestrator mode and Dynamic Workflows were not tried by us
@sundarpichai summarizes what Google released this week: two eighth-generation TPUs, Gemini 4 Carbon, Gemini Robotics 2.5, AlphaGenome 2, the Gemini Decisions API and Gemini 4 Hydrogen (over 7,700 likes). @googledevs says Gemini 4 Hydrogen is in Google Antigravity with an Orchestrator mode, in which Carbon or Argon does the planning so heavy models do not have to carry out every small edit. Antigravity also mentions Dynamic Workflows: with /run-workflow the agent writes a Python script that runs parallel and sequential subagents. According to Google, Google Wallet can now be linked to Gemini through Personal Intelligence.
What this means: a heavy model that plans and a lighter model that executes now appears in several tools. Test such a split on your own tasks and measure cost and error rate. Background on the models is at Gemini 4 Hydrogen and Gemini 4 Carbon.
Sources: @sundarpichai op X (weekoverzicht Google) · @googledevs op X (Orchestrator-modus) · @antigravity op X (Dynamic Workflows) · @Google op X (Wallet in Gemini)
ChatGPT Wallet: agents can make purchases through Dots, first for Pro users in the USConfidence: medium: post by @testingcatalog and a reference to TechCrunch read on X; the OpenAI announcement was not read back
According to @testingcatalog, ChatGPT Wallet is rolling out on Dots for Pro subscribers in the US: agents can pay without real card details and with a security check per purchase. According to our search round TechCrunch also reported the rollout. Reports of an EU rollout are circulating that we could not confirm.
What this means: anyone who lets an agent pay should set beforehand which purchases may go ahead without approval and up to what amount. The same question applies to your own agents: which actions may run without sign-off.
Sources: @testingcatalog op X (ChatGPT Wallet op Dots) · @TechCrunch op X (ChatGPT Wallet)
Agent tooling security: Cursor incident update, an opencode report about OAuth URLs and local MCP serversConfidence: low to medium: the posts by Cursor and Anthropic are official; the opencode, mcp-gateway-registry and localhost reports come from small accounts and were not verified at the source
@cursor_ai published a detailed account of its recent security incident and the remediation, via a link; we did not read the contents. One account reports that opencode up to and including 1.18.32 could run code on Windows/WSL through the OAuth URL of a malicious MCP server (fixed in 1.18.33). Another reports that mcp-gateway-registry 1.0.3 through 1.34.0 had an SSRF that 1.35.0 closes for new registrations. A third warns that an MCP server listening only on localhost can still be reached by websites in the browser when an origin check is missing. The last three reports come from accounts with little reach (0 to 2 likes) and were not checked at the projects themselves. Separately, Anthropic published ten recommendations on its Engineering Blog for designing and reporting safety evaluations (over 1,600 likes).
What this means: if you use opencode on Windows/WSL, check that you are on 1.18.33 or higher. Have local MCP servers check the Origin header and install only servers you trust. Anthropic's recommendations can serve as a framework for your own evaluations before an agent gets write access. Context is in our overview of AI and cybersecurity and our overview of MCP servers and agent tooling.
Sources: @cursor_ai op X (incidentupdate) · @NeuralAlphaAI op X (opencode en OAuth-URL) · @swif_ai op X (mcp-gateway-registry) · @mcpprogrammer op X (lokale MCP-servers) · @AnthropicAI op X (safety-evaluaties)
llama.cpp v1 and Cloudflare Kitesurf: open-source building blocks for local models and browser agentsConfidence: medium: posts read on X; the releases themselves (GitHub, release notes) were not checked
Clement Delangue (Hugging Face) announced llama.cpp v1 together with Georgi Gerganov, the first major stable release of the widely used inference library (over 2,500 likes). @CFchangelog says Cloudflare has open-sourced Kitesurf, a lightweight browser engine for AI agents under Apache 2.0 that runs on Workers. A user points out that a self-hosted instance runs without authentication by default. Cloudflare also put Spend Management in open beta: set thresholds and have activity pause once they are reached.
What this means: a stable v1 makes running locally more predictable for anyone who pins versions; read the release notes before you update. Do not expose a self-hosted Kitesurf to the internet without authentication.
Sources: @ClementDelangue op X (llama.cpp v1) · @CFchangelog op X (Kitesurf) · @CloudflareDev op X (Spend Management)
Unconfirmed: layoffs at Grok Imagine, the scale of the Microsoft outage and an apparently smaller Claude Pro limitConfidence: low: an unconfirmed scoop, an official but vague statement and a user's own calculation
@aatilley reports as a scoop that SpaceX laid off the team behind Grok Imagine (about 60 people); the tool is still live and we found no confirmation from xAI or SpaceX. @Microsoft wrote that it is deeply sorry for an incident that affected customers and that broad recovery is under way; the post names only a technical problem, and reports from third parties circulate about the scale. One user calculates that 84 percent of the Claude Code weekly limit two weeks ago was about 515 million tokens and 51 percent is now about 220 million, mostly on the $20 plan; we found no response from Anthropic. Under viral posts there are many questions to @grok about whether something is true, on unconfirmed news.
What this means: treat this as rumor. Check an outage on the provider's status page, measure your own usage instead of adopting a screenshot, and wait for an official announcement about Grok Imagine.
Sources: @aatilley op X (Grok Imagine-team) · @Microsoft op X (storing en herstel) · @Octoberry op X (Claude Code-limiet)
Primary sources: @ClaudeDevs: mods en thema's in Claude Code · @deepseek_ai: app-functies in de API · @sundarpichai: weekoverzicht Google
