Radar, November 4, 2026: ten signals from X on Claude threads, connectors, MCP and Codex
In short: The standout signal of November 4: the Claude API now keeps the conversation itself with message threads. Alongside it, ten short signals on Claude Code, connectors, MCP security and Codex, each with its source on X. This edition was compiled afterwards, on November 8, 2026, from the X briefings of November 4.
Illustration by Claude (AI), drawn with code: a radar screen with ten signals.
1.Claude API now keeps the conversation itself with message threads
Confidence: high (own announcement)Anthropic put message threads into public beta on the Messages API. The API holds the conversation between turns, so each turn you only send the new message. According to @ClaudeDevs this makes Claude Code faster with long context. If you build your own agent, the main saving is traffic and the code that tracks history.
Source on X: @ClaudeDevs
Context in the briefing: Message threads in the November 4 briefing
Read also: Claude Code news
2.Claude Code in October: mods, parallel sessions and ultrathink per turn
Confidence: high (own announcement)Anthropic summed up the October updates for Claude Code: mods, a desktop app that handles many sessions side by side better, and more insight into usage. The word 'ultrathink' sets effort to xhigh for a single turn. Useful if you only want to spend more of your limit on one hard step.
Source on X: @ClaudeDevs
Context in the briefing: Claude Code in the November 4 briefing
Read also: Claude Code news
3.Claude auto-connects connectors at sign-in
Confidence: high (announcement and documentation)Team and Enterprise organisations can now have connectors connect automatically as soon as someone signs in to Claude. Anthropic talks about plugins on X; the documentation calls it Enterprise-managed auth. The variant through your own identity provider is generally available, the variant with Claude as trusted issuer is in beta and only checks organisation membership.
Source on X: @ClaudeDevs
Context in the briefing: Connectors in the November 4 briefing
Read also: Claude auto-connects connectors · MCP and agent tooling
4.AWS API MCP Server accepted tokens meant for other apps; update to 1.6.0
Confidence: medium (report with little reach)According to a report on X, versions 1.3.0 to 1.6.0 of the AWS API MCP Server accepted OAuth tokens meant for another app. That allowed AWS APIs to be called with the host's permissions. The post has little reach; check the version you run and AWS's own advisory.
Source on X: @hoshino_popopo_
Read also: MCP and agent tooling
5.Coding agents often install outdated MCP servers, research says
Confidence: low (single post, research not read)Glow Labs reports that AI coding agents regularly install MCP servers that are no longer maintained and therefore get no patches. The post drew hardly any attention and we have not read the research. The advice stays simple: do not let an agent add an MCP server unless you check the source and the latest release yourself.
Source on X: @f1tym1
Read also: MCP and agent tooling
6.Vercel checks every push for vulnerabilities
Confidence: high (own announcement)Vercel launched Security Review in public beta: every push gets a vulnerability review, and Vercel Agent can create a fix with one click. For small teams without their own security review that is an extra safety net; a proposed fix is still something to read yourself.
Source on X: @vercel
Read also: Skills and plugins
7.Unity releases an official plugin for Codex
Confidence: high (own announcement)Unity put an official plugin into the OpenAI Codex plugin directory. Codex picks the skills from the plugin that it needs to finish a Unity task. It is an example of software makers shipping their own know-how as skills for coding agents.
Source on X: @unitygames
Read also: Codex news · Skills and plugins
8.Developers keep complaining about the shared ChatGPT and Codex limit
Confidence: medium (user criticism)A widely shared post asks OpenAI to separate the ChatGPT and Codex limits again. Anyone who brainstorms in ChatGPT and builds in Codex now draws from one budget. OpenAI did not respond in this post; with heavy use, keep track of usage per tool.
Source on X: @buildwithrajath
Read also: Codex news
9.Anthropic researches with FlashProve which model produced an answer
Confidence: high (own announcement, research)Anthropic presented FlashProve, research into 'provable inference'. Using cryptography, a developer should be able to verify that an answer really comes from a particular model. That matters if you buy model access through an intermediary and want to be sure no cheaper model was swapped in.
Source on X: @AnthropicAI
Context in the briefing: FlashProve in the November 4 briefing
10.Grok Bot can make phone calls itself, Musk says
Confidence: medium (promotional post; shopping feature unconfirmed)Elon Musk promoted that Grok Bot can make phone calls, for example to find a doctor and book an appointment. According to TestingCatalog a shopping feature with a deal finder and passkeys is also coming. The calling feature was promoted by xAI itself; the shopping feature has not been announced yet.
Source on X: @elonmusk
Context in the briefing: Grok Bot in the November 4 briefing
