Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 13, 2026 · morning edition

Agents are shedding the single vendor account, with Brave selling search to a wallet instead of an API key, Edge0 putting a 145B model on one NVIDIA GPU plus system RAM, and gateway repos routing Claude Code to other providers, on a morning when a Claude sign-in incident showed what one account costs.

Brave Search APIx402Edge0Claude CodeCheckpoint Forensicsagent-paymentslocal-aiclaude-codeagent-infrastructurefrontier-models

Trending AI Briefing: Friday, November 13, 2026 (morning ET)

Friday's signal is quieter at the big labs and louder at the edges. The pattern: an agent no longer needs one vendor account to do its job. Brave now sells search to any agent that can pay a few tenths of a cent in USDC, with no key and no signup. Edge0 shipped a 145B model its README pitches at a single NVIDIA card, with system RAM holding the overflow. Two of today's busiest repos are gateways that point Claude Code and Codex at other providers. Then, before dawn on the East Coast, Anthropic opened an incident for elevated errors on Claude and Claude Code, its only update naming a sign-in problem. That is two companies and two community projects, not a lab-wide shift, but they all push the same way.

What's hottest in AI news right now

Brave made its Search API payable per request by agents, with no account or API key. The launch post, dated November 12, puts LLM Context (Brave's search endpoint built to feed agents) at $0.005 a query, $5 per 1,000, and says web, news, image, video and local search are available at the same rate. The flow is the open x402 protocol: an unpaid request gets an HTTP 402 with the price, the agent signs a USDC payment on Base and retries, and results come back with a receipt. Brave's llms.txt adds the part builders will like: "A failed search is never charged," because the payment settles only after the search runs. The honest catch: Base USDC is the only rail today, gas is covered "for a limited time" with no end date, and neither the docs nor the rate-limiting guide say what rate limits apply to wallet-paid calls. Brave blog · Agent payments guide

Edge0 released Edge0-145B and Edge0-120B, sparse models aimed at one consumer GPU. The announcement is dated November 12; the repo's release note is dated November 13. Per the README, both are latent mixture-of-experts models (each token wakes 10 of 512 experts) derived from Qwen3.8-Flash-Next; the blog does not name a base model and quotes its own, lower figures, such as 80+ tokens per second. The README reports 121.2 tokens per second of agent decode for the 145B Q3_K build on an RTX 5090 D with 24 GB, and a 145B score of 86.5 on Terminal-Bench 2.1 against 78.2 for Claude Opus 4.6 (Max). Those are Edge0's own runs, against an older Claude. The bigger catch is in the README itself: the 145B and 120B models run only on Linux or Windows with an NVIDIA GPU and CUDA 13, and the M2 Ultra speed row "does not establish macOS release availability," even though the blog pitches Apple M-series laptops. The weights also carry the Edge0 Community Model License, which requires written permission for paid model services and commercial coding assistants. Edge0 blog · README · HN

Claude Code 2.1.323 added Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. npm published 2.1.323 on November 12 at 20:29 UTC, and the changelog points to a docs page that covers turning it on, or off for a whole organization. The same release lets plugins declare a supported Claude Code range through engines["claude-code"] in plugin.json, so out-of-range plugins are not installed or loaded, and adds a 30-second heartbeat while long hooks run in claude -p. The catch: npm's stable tag still points at 2.1.311, so teams on the stable channel do not have it, and the Code Mode docs page did not load for this briefing, so setting names are not confirmed here. Changelog · npm

A forensic write-up says six Chinese open-weight models were initialized on export-controlled Blackwell GPUs. Will DePue's Checkpoint Forensics, published and revised November 12, reads the random-initialization pattern PyTorch leaves in untouched embedding rows, which depends on the GPU's count of streaming multiprocessors (132 on Hopper, 148 or 152 on Blackwell parts). It names DeepSeek V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax M3 and Ling-3.1-flash, calls the V4-Flash result "conclusive" while rating the other five "strong evidence," and published code. The authors' own limits matter: it shows where starting weights were generated, not where training ran or where the chips sat, and it does not show an export violation. They also say of their GPU-count estimates, "We are not confident in these estimates." Report · HN

Anthropic opened an incident for elevated errors on claude.ai, Claude Code and platform.claude.com. The status page posted it at 09:58 UTC on November 13, and its only update, still "investigating" when this briefing read it, named users authenticating to Claude.ai and Claude Code when this briefing read it, with Claude Cowork also affected. No cause or fix time was posted. Hacker News had two threads on it within about 20 minutes. This is short-lived news, but it lands on the same morning as three stories about agents working without one account. Status · HN

New tools and features worth actually trying

Brave's paid search proxy through purl. Fund a dedicated Base wallet, install Stripe's purl (brew install stripe/purl/purl), and run Brave's documented command purl --max-amount 5000 'https://search.agent.s.brave.app/res/v1/web/search?q=rust'; the cap of 5000 base units is exactly one $0.005 search per Brave's llms.txt (purl's own README does not list the flag, so check purl help). Honest tradeoff: you now hold crypto for your agent, receipts live on-chain rather than in a dashboard, and paid-call rate limits are undocumented.

Code Mode in Claude Code 2.1.323. Worth a test on any workflow that chains many small tool calls, since the model writes one script instead of a long round-trip sequence. Honest tradeoff: it is experimental, it ships on the latest and next channels, not stable, and generated code calling your tools deserves the same review as any code that touches your machine.

zigzag's built-in MCP server. The Show HN embroidery editor exposes about 30 tools (add_svg, render_preview, check_design, send_to_skitch) at 127.0.0.1:7878/mcp, a clean example of a desktop app made agent-editable without a cloud API. Honest tradeoff: only the Brother Skitch PP1 is tested on real hardware, and the machine protocol was reverse-engineered, so the README warns against firmware updates.

Claude as an Apple Foundation Models provider. Anthropic's ClaudeForFoundationModels Swift package lets an app's LanguageModelSession call Claude through the same API Apple uses for its on-device model (framework page). Honest tradeoff: it is a beta that needs the 27-series OS and Xcode betas, App Attest needs a physical device, and maintenance is best-effort with no outside contributions. This is not new this week; Apple's framework page reached Hacker News on November 12.

Trending AI repos on GitHub today

Read from Trendshift's daily board at about 07:15 ET; its figures are momentum scores, so star counts below come from cache-busted shields reads, and release feeds were rate-limited for most repos.

  • vercel-labs/ui (#1): Vercel's beta "UI SDK," reusable behaviors (focus, selection, popups, drag and drop, ARIA) that agents and people compose into components for React, Vue, Svelte, Solid or Web Components. Why now: new this week and ships an agent skill (npx skills add vercel-labs/ui --skill ui-sdk). Apache-2.0, about 2.9k stars, release unverified. Caveat: beta terms, APIs may change before GA.
  • Edge0-AI/Edge0 (#6): the inference framework behind Edge0-145B. Why now: the 145B and 120B release above. Framework Apache-2.0, weights under the Edge0 Community Model License, about 11k stars. Caveat: the new models need an NVIDIA GPU with CUDA 13; benchmarks are self-run.
  • h4ckf0r0day/obscura (#7): a Rust headless browser with V8 and the Chrome DevTools Protocol for scraping and agent automation. Apache-2.0, about 35k stars. Caveat: its speed comparisons against Chrome are self-reported, stealth features carry terms-of-service risk it does not discuss, and it carries sponsor links plus a paid cloud waitlist.
  • CopilotKit/OpenBot (#12): self-hosted AI "coworkers," each with its own browser, files and tools, with actions gated before and logged after. MIT template, about 12k stars. Caveat: CopilotKit Intelligence is licensed separately (a 30-day renewable local license), and the example config makes everyone who can reach it an admin.
  • Alishahryar1/free-claude-code (#15): a local proxy that points Claude Code, Codex and other harnesses at many providers, with fallback across keys. AGPL-3.0, about 70k stars, v6.19.11 on November 12. Caveat: the "1.3B+ free tokens" claim has no breakdown, and routing personal coding plans through a proxy may break those plans' terms.
  • simpsoka/zigzag (#18): the MCP-enabled embroidery editor above, more than 560 points on Hacker News. Apache-2.0 LICENSE file, about 450 stars. Caveat: one tested machine.
  • yetone/magpie (#20): a menu bar gateway that runs Codex on DeepSeek or Claude Code on Kimi and fails over when a quota runs out. MIT, about 36k stars. Caveat: sponsor links route through click-counting redirects, and privacy masking is off by default.
  • anthropics/ClaudeForFoundationModels (not on Trendshift): Claude as a provider for Apple's Foundation Models framework. Apache-2.0, about 360 stars. Caveat: beta OS and Xcode required.

What actually matters from today's signal

Track the decoupling. An agent's identity used to be one API key at one vendor, and every capability hung off it. Today it can pay per call from a wallet (Brave), run its model on a card under the desk (Edge0), and swap its backend from a menu bar (magpie, free-claude-code). For builders, the highest-signal areas are per-request payment rails for tools, local inference sized for one GPU, gateways that sit between a harness and its model, and plugin version pinning like the new engines field, which keeps a fleet of agents from loading a plugin it cannot run.

The counter-signal is that each escape hatch moves risk somewhere less watched. A wallet is a spending key; a gateway sees every prompt and every credential you route through it; a "local" model ships under a license that forbids the commercial assistant you might build with it. The Checkpoint Forensics write-up adds one more: open weights carry a provenance trail that most teams never read. Decoupling is good engineering. Do it with a budget cap, a license check and a fallback you have actually tested, not because the main account went down at 5 a.m.

Source access notes: Read November 13 between about 07:15 and 07:40 ET. openai.com/news failed to load; Anthropic news had only the November 10 IPO roadshow post in range; blog.google AI and developer index pages served stale 2024 and 2025 snapshots, so Google items came from direct post URLs (Gemini 4 Hydrogen, already covered November 12, was skipped); the Code Mode docs page and the SpaceBattles thread alleging Humble Bundle security failures from vibecoding failed to load, so the latter is not reported. Microsoft Agent Framework, LangChain, Mistral and xAI had nothing dated November 11 to 13. Hacker News came from the Algolia API (stories since about 00:15 UTC November 12). The workspace shell's outbound fetches were blocked except GitHub raw files and npm, so pages were read through a fetch agent. GitHub release feeds rate-limited on 11 of 12 repos; only free-claude-code's release is confirmed. Trendshift's number strings run star counts and scores together, so shields figures were used instead. Adversarial pass (Sonnet 5.5) checked dates, numbers and attributions; it caught an unsupported obscura benchmark mismatch (dropped), an Edge0 base-model claim only the README makes (now attributed, with the blog's lower figures noted), an overstated status-incident title and a Code Mode "administrators" wording the changelog does not use (both fixed), a purl flag only Brave documents (attributed), a misattributed Hacker News item for Apple's framework (fixed), and a missing contrast between "conclusive" and "strong evidence" in Checkpoint Forensics (added). Star counts could not be re-checked by the pass (rate limits) and remain approximate shields reads.