AI News of the Week: MCP, Grok 4.5, GitHub Models

Updated July 30, 2026. Review period: July 23–30.

The biggest AI news of the week was not a new model, but the maturation of agent infrastructure. MCP got its biggest update and became a stateless protocolGitHub allowed skills and MCP to be connected to code review, and companies united around an open security stack. At the same time, Grok 4.5 arrived in Copilot, GitHub Models is shutting down for good, and NVIDIA showed local multimodal agents running on a compact Jetson.

The practical takeaway: choosing one “best neural network” is no longer enough. Teams need separate solutions for models, tools, access rights, logging, and the execution environment. This week, the market made each of these layers significantly more independent.

In brief: check your migration off GitHub Models today, pin the MCP version, restrict the agent’s tools to the minimum required permissions, and decide in advance which new models can be enabled automatically for employees.

Contents

• What happened this week
• MCP 2026-07-28: agents got a protocol built for scale
• Skills and MCP entered regular code review
• Grok 4.5 appeared in GitHub Copilot
• GitHub Models is shutting down: what needs to be migrated
• Agent security is becoming its own stack
• Local AI is moving out of the server room
• What to do in the next 48 hours
• FAQ
• Bottom line

What happened this week

Date	Event	What changes	Who should care
July 23–28	GitHub and the MCP Project confirmed MCP support 2026-07-28	the protocol becomes stateless, with extensions and formal compatibility tests	platform teams and MCP developers
July 27	the Open Secure AI Alliance was announced	agent security is taking shape as an open, multi-vendor stack	CISOs, architects, AI integrators
July 28	Grok 4.5 started rolling out in GitHub Copilot	another reasoning model is available in a familiar developer interface	developers and Copilot admins
July 28	NVIDIA showed local voice-and-vision scenarios on Jetson	an agent can work without the cloud, API keys, or the internet	robotics, manufacturing, edge teams
July 29	skills and MCP in Copilot code review reached GA	code review now gets internal standards and external context	engineering leaders
July 30	GitHub Models is fully shutting down	the playground, catalog, inference API, and BYOK are no longer available	all GitHub Models users

The events and dates in the table are Measured based on primary sources. Practical significance is Estimated editorial analysis from AI Dawn.

MCP 2026-07-28: agents got a protocol built for scale

MCP, or Model Context Protocol, is an open protocol through which an AI agent connects tools and data sources. The 2026-07-28 specification was the biggest MCP overhaul since launch.

The main change is a stateless core, meaning there is no longer a required session at the protocol level. Previously, a remote MCP server could require initialize, a session ID, sticky routing, and shared state storage. Now a request can be sent to any instance behind a standard load balancer. Business-process state, if needed, is passed explicitly — for example, via basket_id or browser_id.

This does not mean all applications should forget context. The transport layer has changed: state is no longer hidden in the connection. For operations, that brings four advantages:

• MCP servers are easier to scale horizontally;
• calls are easier to trace and route;
• the tool list can be cached;
• less infrastructure is needed just to support a session.

Along with the new core, extensions were added for MCP Apps and long-running Tasks, authorization was strengthened, formal compatibility tests and a deprecation policy were introduced. GitHub moved its MCP Server ahead of time to the new schema: it removed Redis sessions and database reads on every call.

There is a cost to the transition: the version includes breaking changes. A team with its own MCP server needs to pin supported client and SDK versions, pass the conformance suite, and separately test older consumers.

Skills and MCP entered regular code review

On July 29, GitHub made agent skills and MCP in Copilot code review generally available for Pro, Pro+, Business, and Enterprise plans.

A skill is a local package of instructions and context in SKILL.md. A team can place it in .github/skills, so code review takes internal standards into account: architecture rules, migration requirements, API conventions, or a security checklist. MCP servers add data from the issue tracker, documentation, or the service catalog.

This matters more than yet another chat interface. AI review is no longer limited to the model’s general knowledge: it can match a pull request against repository-specific rules and the company’s current context.

GitHub has limited MCP calls in code review to read-only mode and added attribution for comments created with a skill or MCP. That is a sensible boundary, but it does not remove the need to verify:

1. Which documents and services does the agent see?
2. Can the issue content or documentation contain prompt injection?
3. Which secrets are used for the connection?
4. Are requests and responses from the external provider logged?
5. Can a developer tell which source influenced the comment?

Separately, GitHub is changing model management. Starting August 26, new GA models in Business and Enterprise will inherit the global availability policy. By default, the policy is enabled. If a company wants to approve each model manually, the administrator needs to turn it off in advance.

Grok 4.5 is now available in GitHub Copilot

On July 28, GitHub began rolling out Grok 4.5 gradually in VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, JetBrains, Xcode, and Eclipse.

According to GitHub’s specifications, the model accepts text and images, supports up to 500,000 context tokens and three levels of reasoning effort. GitHub reports strong internal results on terminal tasks, parallel tool calls, and urgent multi-step scenarios. These are vendor-provided data, not an independent AI benchmark.

The practical value of the release is lower switching costs between models. A developer does not need to integrate a separate API to compare Grok 4.5 with other options on the same repository.

For Business and Enterprise, access to Grok 4.5 still needs to be explicitly enabled by an administrator. Before enabling it, compare:

• the share of tasks completed without a rerun;
• the number of faulty changes in a pull request;
• time to an accepted result;
• AI credit usage per completed task;
• code retention and transfer policies.

GitHub Models is shutting down: what needs to be migrated

On July 30, GitHub is fully shutting down GitHub Models. The playground, model catalog, inference API, and BYOK are being retired. The change affects all customers, including projects with active usage.

If an app calls models.github.ai, simply replacing the model in the settings may not be enough. Check five things:

What to check	Why it breaks during migration
base URL and authentication	the new provider uses a different endpoint and key
model IDs	names and versions do not match
streaming and tool calls	event and error formats differ
limits and timeouts	latency, quotas, and retries change
data policy	the new provider may store requests differently

GitHub recommends Microsoft Foundry for the model catalog and Copilot for AI scenarios inside GitHub. Those are not the only options. From an architecture and security standpoint, it is safer to put the model call behind your own adapter, keep a set of contract tests, and avoid tying business logic to a single response format.

If GitHub Models is still in production, this is no longer a planned migration but an availability incident: you need to switch traffic, check for degradation, and only then optimize quality and cost.

Agent security is becoming its own stack

On July 27, NVIDIA and dozens of participants including GitHub, Microsoft, Hugging Face, IBM, Mistral, LangChain, Cloudflare, and major cybersecurity companies announced the Open Secure AI Alliance.

The stated goal is to build open models, harnesses, and tools to protect software and agents. The initiative includes different layers: zero-trust identity for services, secure weight formats, signed patches, multi-model scanning, and agent behavior tracing.

The most useful point in the announcement is that an agent is not just an LLM. Security is defined by the entire system:

• agent and user identity;
• tool permissions;
• environment isolation;
• action limits and confirmations;
• logs and tracing;
• tests for malicious context;
• the vulnerability remediation process.

The alliance should not yet be taken as proof that these tools are mature or interoperable. It is an announced program and a set of participant contributions. But the direction matters: the debate over an open or closed model gives way to a more practical question—can the full path from prompt to action be verified and controlled?

Local AI is moving out of the server room

In the same period, NVIDIA showed local agent workflows on Jetson. Jetson Orin Nano Super is listed with 67 TOPS, and AGX Orin with 275 TOPS. TOPS is trillions of operations per second; the figure helps compare device classes, but by itself does not guarantee the speed of a specific model.

A practical example is the Reachy Mini voice and visual assistant: recognition, vision, and responses all run locally with GPU acceleration, without cloud services, API keys, or internet access at runtime. NVIDIA also released Jetson Device Skills and BSP Skills so coding agents can help configure and deploy edge systems.

Edge AI, or on-device AI, isn’t necessary for every app. It becomes interesting when latency, offline operation, the cost of a constant data stream, or a ban on sending video and audio to the cloud are critical. The tradeoff is limited memory, fleet-wide updates, and the team’s responsibility to optimize the model.

What to do in 48 hours

This week’s AI news turns into a quick infrastructure audit.

1. Find GitHub Models. Search for models.github.ai, the token models:read and legacy model names. Switch production traffic.
2. Record your MCP versions. Note the protocol, client, and SDK versions, then run compatibility tests before switching to 2026-07-28.
3. Check agent permissions. For each tool, define read, write, approve, and deny. Leave code review read-only.
4. Define the model policy. Decide whether new GA models can be enabled automatically. For a locked-down environment, disable inheritance until August 26.
5. Compare on completed task value. Test Grok 4.5 and other models based on the cost of an accepted result, not token cost or a single benchmark.
6. Pick one edge pilot. If you have a camera, robot, or production system, test a local scenario with measurable latency and offline mode.

FAQ

What was the biggest AI news of the week through July 30, 2026?

The biggest platform release was MCP 2026-07-28: the protocol became stateless and gained extensions, stronger authorization, and formal compatibility tests. For GitHub users, the most urgent event is the complete shutdown of GitHub Models.

What does stateless MCP mean?

The protocol no longer requires a hidden session between the client and server. Requests can be distributed across instances by a standard load balancer, and the necessary business state can be passed explicitly in tool arguments.

Do we need to update the MCP server urgently?

Not blindly. The new version includes breaking changes. First lock the clients and SDKs, verify backward compatibility, and run conformance tests.

What can replace GitHub Models?

GitHub recommends Microsoft Foundry for access to the model catalog and Copilot for working inside GitHub. You can choose another provider if your app uses an adapter and contract tests, and your data, tool calls, and streaming requirements have been verified.

Should we enable Grok 4.5 for the whole team?

Start with a limited test on real repositories. Compare time to an accepted change, errors, reruns, credit usage, and code-handling rules. In Business and Enterprise, the model requires admin enablement.

Will local AI replace cloud models?

No. Edge AI wins on offline capability, latency, and privacy, but it is limited by device resources and needs its own maintenance. More often, it becomes a separate layer rather than a full replacement for the cloud.

Bottom line

From July 23 to 30, the AI market moved from demos to infrastructure. MCP made it easier to scale tools. Skills and external context became part of standard code review. Administrators got new model-management tasks. Agent security emerged as a separate open stack, and local systems moved closer to real devices.

So the best question of the week is not “which model won?” but “can we replace the model, limit its permissions, verify the source of context, and restore service after a vendor outage?” If you can answer all four, your agentic system is ready to mature.