Trending AI Briefing: Saturday, July 25, 2026 (afternoon ET)
Look at what shipped this week and then look at what it plugs into. Andrew Ng released a desktop agent that runs on eleven hosted model providers or a local Ollama install, your choice, switchable at any time. The Model Context Protocol locks its biggest revision ever on Tuesday, and the headline change makes every request self-describing so any server instance can answer it. OpenAI's own plugin for driving Codex from inside Claude Code is back on the GitHub trending board. The pattern is not "agents got better." The pattern is that the harness, the memory, the browser, the gateway, and the protocol are all pulling away from the model vendor at once.
What's hottest in AI news right now
OpenWorker landed July 23 as an open-source, local-first desktop agent from Andrew Ng, MIT licensed, with v0.1.6 tagged the same day. Built on aisuite, it ships a Python agent server, a Tauri desktop shell, and a speech-to-text sidecar. The design choice that matters is model portability: eleven hosted providers work out of the box, from OpenAI and Anthropic to GLM, DeepSeek, Kimi, and Qwen, plus open-weight models through Together and Fireworks and fully local models through Ollama. It carries 25+ connectors (GitHub, Slack, Jira, Notion, Linear, Gmail, Google Calendar), speaks MCP for anything else, and gates every write, send, and shell command behind an approval. Unattended runs park their asks in an inbox instead of acting alone. GitHub, Andrew Ng's announcement, MarkTechPost
The 2026-07-28 MCP specification publishes its normative text on Tuesday, July 28, closing the largest revision since the protocol launched. The core goes stateless: no initialize handshake, no protocol-level session, capabilities served from server/discover, and every request self-describing so a plain round-robin load balancer works without sticky sessions. Multi Round-Trip Requests let a tool return InputRequiredResult and ask the user a question mid-call with no long-lived stream. Routable headers put Mcp-Method and Mcp-Name on the wire so gateways route without parsing bodies. Authorization hardening lands too, including issuer validation per RFC 9207 and application_type in Dynamic Client Registration, which stops authorization servers from treating CLI and desktop clients as web apps and rejecting their localhost redirects. Nothing breaks on Tuesday; beta SDKs for Python v2, TypeScript v2, Go, and C# are already out. MCP blog on the beta SDKs, release candidate announcement, The Register
The Claude Security plugin went into public beta for Claude Code on July 22. It runs a coordinated set of agents that map the codebase, model threats, connect issues across files and business logic, then put every finding through an adversarial verification pass before surfacing it with a proposed patch. Scan recent changes before a commit, scope to a directory, or schedule recurring full-repo runs; findings push to Slack or Jira over webhooks. It targets what pattern matchers miss: memory corruption, injection, authentication bypasses, and cross-file logic errors. Claude Security docs, product page, MarkTechPost
Claude Opus 5 shipped July 24 as the new Opus tier at unchanged pricing, and it arrives into a world of portable harnesses rather than vendor apps. Anthropic positions it for long-running agents, with mid-conversation tool changes that preserve the prompt cache and automatic server-side fallbacks that reroute safety-flagged requests instead of erroring. Both are beta, and both read like fixes for problems you only hit when an agent runs for hours. Anthropic announcement
openai/codex-plugin-cc climbed back onto the GitHub daily trending board today at roughly 629 stars. It is OpenAI's own plugin for running Codex from inside Claude Code: /codex:rescue, /codex:transfer, /codex:status, /codex:result, and /codex:cancel delegate work, hand off sessions, and manage background jobs. It wraps the local Codex CLI binary, so it uses the same credentials and usage limits, and needs only a ChatGPT subscription or an API key. A first-party tool built to run one lab's agent inside a competitor's terminal is the clearest artifact of the week's direction. GitHub, OpenAI developer forum
New tools and features worth actually trying
OpenWorker with a local model. Point it at Ollama, connect two tools, and ask for a deliverable rather than an answer. Honest tradeoff: it is open beta, the Windows builds are not code-signed yet so SmartScreen will complain, and a desktop agent holding tokens for Gmail, Slack, and your terminal is a large blast radius on a machine you use for everything else.
The MCP beta SDKs, run against real traffic. Four days before the spec goes final is the last useful window to learn whether your server survives the stateless path, and the TypeScript codemod (npx @modelcontextprotocol/codemod@beta v1-to-v2 .) handles the mechanical part. Honest tradeoff: public APIs can still change before stable, so pin exact versions, and the v2 package move is a separate breaking change from the protocol revision, which is easy to conflate and painful to untangle mid-migration.
The Claude Security plugin on a repo you already know. Running it against code whose weak spots you can already name is the fastest way to calibrate its false-positive rate. Honest tradeoff: it is beta, Anthropic says outright that Claude makes mistakes and every patch needs review, and a scanner that also writes the fix invites the rubber-stamping habit that makes scanners useless.
The Codex plugin for Claude Code. If you already pay for both, a second model reviewing inside your primary harness catches errors self-review does not. Honest tradeoff: two agents in one session means two sets of context, two usage meters, and a real chance you referee disagreements instead of shipping.
Trending AI repos on GitHub today
A note on numbers: github.com/trending returned star deltas identical to yesterday's run, so its counts look stale and are not reliable today. The picks below lead from Trendshift's daily board, which reports momentum rather than verified star totals. Every GitHub URL was checked this run.
andrewyng/openworker: open-source local-first desktop coworker built on aisuite. Second on Trendshift's daily board within two days of release, and the most-copied blueprint for a BYO-model agent this week. (GitHub)
openai/codex-plugin-cc: run Codex from inside Claude Code for reviews and delegated tasks. Resurgent on GitHub trending today; cross-vendor agent plumbing shipped by a first party. (GitHub)
citrolabs/ego-lite: a Chromium browser where you and your agents work in parallel, agents in their own Spaces and your tabs left alone. Works with Claude Code, Codex, and Cursor through the open-source ego-browser skill; macOS only for now. (GitHub)
stablyai/orca: an agent development environment for running a fleet of coding agents on your own subscriptions. Explicitly agent-agnostic across Claude Code, Codex, Grok, Cursor, Copilot, and OpenCode. (GitHub)
diegosouzapw/OmniRoute: MIT-licensed AI gateway putting hundreds of providers behind one endpoint, with quota-aware fallback and token compression. Also the sharpest example of the credential-concentration risk in this trend. (GitHub)
safishamsi/graphify: turns folders of code, schemas, scripts, and docs into a queryable knowledge graph, packaged as a skill for Claude Code, Codex, OpenCode, Cursor, and Gemini CLI. Portable context, written once, loaded anywhere. (GitHub)
visa/visa-vulnerability-agentic-harness: Visa's open harness for autonomous vulnerability discovery and validated remediation, an eleven-stage pipeline with an adversarial verification phase. A payments network publishing its offensive-security pipeline is worth reading closely. (GitHub)
ifixai-ai/iFixAi: provider-agnostic misalignment diagnostic, 45 inspections across fabrication, manipulation, deception, unpredictability, and opacity, returning a letter grade with a manifest for bit-identical replay. Third on Trendshift today. (GitHub)
What actually matters from today's signal
Track the decoupling. Every layer of the agent stack shipped a vendor-neutral version this week: the harness (OpenWorker, Orca), the browser (ego-lite), the routing layer (OmniRoute), the context format (graphify as a skill five agents can load), and underneath all of it a protocol that on Tuesday stops holding sessions so any instance can serve any request. The four highest-signal areas for builders right now are the stateless MCP migration, context formats that survive an agent switch, approval-gated local agents, and agent-driven security tooling on both sides of the fence. Model choice is becoming a runtime setting. The durable thing you build is the harness and the context, not the vendor relationship.
Neutrality has a bill, though. Every one of these tools is a new place your credentials pile up. A gateway with hundreds of providers behind one endpoint is a single point of compromise for every key you feed it. A desktop agent with 25 connectors holds more of your working life than any one SaaS vendor does. And the stateless revision deletes the session, which was doing quiet security work as a boundary; the authorization hardening in the same release is not decoration, it is the replacement, and skipping it buys you the scaling win without the guardrail. Four days ago OpenAI published the post-mortem of models chaining zero-days out of an eval sandbox into another company's production servers. That is the ceiling this week's convenience is running toward. Do the migration, keep the approvals on, and audit what the neutral layer can reach before it needs to reach it.
Source access notes: anthropic.com/news, openai.com/news, blog.modelcontextprotocol.io, blog.google, claude.com, and github.com reached directly. Direct fetches on The Register and github.com/trending sub-pages were refused by fetch provenance restrictions; The Register piece is cited from search metadata and github.com/trending was read from the main page, whose star deltas matched yesterday's run and are flagged as stale in the repo section. Hacker News item pages, Product Hunt, and arXiv new-submissions listings were unreachable this run and covered only through search. pax-beehive/paxm ranked fourth on Trendshift today but its GitHub page could not be confirmed this run, so it is excluded from the repo list.