AI Weekly #8 — Sep 14–20, 2026
September 14–20, 2026
MCP left the code editor this week: Google shipped it as smart-home control, and Nature published a method that turns research papers into callable servers.
Two things happened to the Model Context Protocol this week, and neither of them happened in a code editor.
On Wednesday Google opened Home MCP in early access: any MCP-capable agent can now list the structures in a house, read live device state, issue commands and query the event history behind a Nest doorbell or a Matter bulb. The same day, Nature published Paper2Agent, a Stanford method that takes a research paper's manuscript, code and tutorials and emits an MCP server a chat agent can call. It converted 74 of 100 computational biology papers without a human touching them.
A protocol that a year ago was mostly a way to hand a coding assistant your filesystem is now the interface between an agent and a thermostat, and between an agent and a peer-reviewed method. That is a different kind of milestone from a version bump, and it is worth more attention than either story got.
The socket generalised
Our MCP catalogue holds 293 servers, and the shape of what arrives has been shifting for months: less filesystem-and-git, more vendor-hosted endpoints sitting in front of something that already has customers, permissions and an audit requirement. Google Home fits that pattern exactly. It needs a Premium Advanced subscription, a Google Cloud project and OAuth; it rate-limits; and it refuses outright to unlock doors. The interesting engineering is not the five tools it exposes. It is the list of things the tools will not do — which is what you would expect from a team that has thought about what a confused agent with your front door in its tool schema looks like. The vendor-hosted servers arriving in the catalogue read the same way: Box's remote server is governed access to files a company already has permissions on, not a new capability.
Paper2Agent is the stranger of the two, and the more consequential. The pipeline identifies a paper's codebase, builds its environment, finds and executes its tutorials, extracts the working steps as tools and assembles them into a server. Across 100 papers it proposed 599 tools and 593 passed validation; the resulting agents scored 91.2% on 300 questions derived from the tutorials. The AlphaGenome agent managed 98.7% on tutorial queries and 100% on novel ones, against 82.7% for a general model pointed at the repository. Three chained paper-agents were used to nominate GPR137 as a causal gene for psoriasis, which was then validated experimentally.
The unit of reuse in computational science has been the paper, which you read, and the repository, which you fight. Making the artifact callable is a bigger claim than "AI reads papers now". It is a claim about what gets published.
Hold onto the 26 that failed, though. A quarter of a curated sample of biology papers could not be made to run by an automated pipeline, and that number is the honest headline — it is a reproducibility measurement wearing a tooling paper's clothes.
Who checks the work
On Friday Anthropic said it will pay Accenture's Faculty arm to embed evaluators inside the company with access comparable to an employee's: watching models take shape in training, reading the deployment decisions, talking to staff. Both sides expect to commit at least $1 billion each over five years. It is the concrete follow-through on Dario Amodei's pacing essay the previous weekend, and it arrives with an admission worth quoting: Anthropic funds the work directly because funding structures for independent evaluation do not yet exist.
That admission is the whole problem in a sentence. An auditor paid by the audited is the arrangement every other regulated industry spent a century learning to distrust, and saying so plainly is better than not saying it. Anthropic notes the deal is non-exclusive, that Accenture will work with other developers, and that METR and other nonprofits are in discussion about piloting the same access on their own money. The version that matters is the last one.
The bill arrives itemised
Two unrelated items priced the substrate underneath all of this.
The House passed the Ratepayer Protection Act 417–3 on Wednesday. It requires state utility commissions to consider large-load standards for data centres drawing 100 MW or more at a single site, so those sites cover the full incremental cost of serving them rather than spreading it across everyone else's bill. It still has to clear the Senate. But 417–3 is not a partisan vote; it is a vote that has already been taken in people's electricity bills, and the legislature is catching up to it.
Cornelis Networks, meanwhile, raised $205 million and published the number from the other side of the meter: in a modelled 100,000-GPU system, roughly half of GPU hours are spent waiting for data — about $1.68 billion a year of idle capacity and 500 GWh of power. Treat a vendor's model of the problem its product solves with the usual scepticism. The direction is still right, and it is the same argument the ratepayer bill is making from the grid end.
By the numbers
The directory took in 38 entries this week — 13 apps, 18 agent skills and 7 MCP servers — bringing the totals to 516 apps, 274 skills and 293 MCP servers. Intake nearly doubled on last week's 21.
Nine of those 38 turned out to have shipped something upstream during the week once we opened the sources. The skills column is the stark one: all 18 skills catalogued this week came from vendor-maintained collections that already existed — Microsoft's Power Platform and Azure plugin marketplaces, Datadog's agent skills, Google Maps Platform, the Dart team's, Anthropic's own — with first commits we could see running from April to August 2026. Exactly one skill, Playwright CLI, cut a release inside the window.
That is not a complaint about the entries; they are real, catalogued and searchable. It is a fact about what "new" means in a category that has stopped launching and started accreting. The MCP servers behaved the same way: Box's dates to September 2025 and Meta's ads connector to April 2026, whatever week we happened to index them. Last week's issue made the same observation about a smaller sample, and it has held.
What we added
Nine entries below shipped upstream between 14 and 20 September, verified against npm publish times, GitHub releases and launch threads rather than the date we filed them.
The week in AI
Google opens the smart home to any MCP agent
Google put its Home MCP server into early access on 16 September, letting any MCP-capable agent discover the structures in a home, enumerate devices and their command schemas, read live trait state, run control actions and query historical events. It exposes five tools — list_homes, list_home_resources, list_home_states, run_home_actions and list_home_history — behind OAuth, a Google Cloud project with the Home API enabled and a Google Home Premium Advanced subscription. Google rate-limits the server and blocks sensitive actions such as unlocking doors.
Why it matters: MCP has been a developer-tools protocol. This is the first time a mass-market consumer platform has used it as the front door to physical hardware, and the refusal list matters more than the tool list.
Nature publishes a pipeline that turns a paper into an MCP server
Paper2Agent, from a Stanford-led group, was published in Nature on 16 September. It identifies a paper's codebase, builds its environment, finds and runs its tutorials, extracts the working steps as tools and assembles them into an MCP server a chat agent can call. Across 100 computational biology papers it converted 74 without human intervention, and 593 of 599 proposed tools passed validation. The resulting agents scored 91.2% on 300 tutorial-derived questions, and three chained paper-agents were used to nominate GPR137 as a causal gene for psoriasis.
Why it matters: It reframes the publishable artifact from something you read to something you call. The 26 papers that would not convert are the more useful number: an automated reproducibility check, reported as a tooling result.
Apple ships Siri AI in beta, built partly on Google's models
Apple began rolling out Siri AI on 14 September across iOS 27, iPadOS 27, macOS 27, watchOS 27 and visionOS 27. Apple says the assistant runs on the next generation of Apple Foundation Models, "custom-built in collaboration with Google and its Gemini models", splitting work between on-device models and Private Cloud Compute. It launches in English, with French, Japanese, Korean, Portuguese and Spanish stated for next month, and requires an iPhone 16 or later, or an iPhone 15 Pro. It is not available in the EU at launch, and remains unavailable in China pending regulatory approval.
Why it matters: Apple spent two years insisting the assistant would be its own. Shipping it on models co-built with Google, and shipping it everywhere except the EU and China, says more about the economics and the regulatory map than any benchmark.
Anthropic will pay Accenture $1bn-scale to audit it from the inside
Anthropic announced on 18 September that Accenture's Faculty arm will embed independent evaluators inside the company with access comparable to an employee's — observing models during training, reading the decisions that govern how they are built and deployed, and speaking directly to staff. The work covers evaluation, red-teaming, alignment assessment and safeguard testing. Both companies expect to invest at least $1 billion each in the capacity over five years. Anthropic funds Accenture's work directly, stating that funding structures for independent evaluation do not yet exist, and says the arrangement is non-exclusive.
Why it matters: It is the first concrete follow-through on the industry pacing argument. It is also an auditor paid by the audited, which Anthropic concedes. The pilots it says METR and other nonprofits may run on their own funding are the version that would settle the objection.
House votes 417–3 to make big data centres pay their own grid costs
The House passed H.R. 9340, the Ratepayer Protection Act, 417 to 3 on 16 September. The bill requires state public utility commissions to consider large-load standards for non-residential customers running data storage and computation at a peak demand of 100 MW or more at a single site or campus, so those customers cover the full incremental cost of serving their load rather than passing it to other ratepayers. It was led by Representative Gabe Evans and now goes to the Senate.
Why it matters: A 417–3 margin on anything touching AI infrastructure is the signal. Electricity prices are the point at which the buildout stops being an industry story and becomes a constituent one.
Anthropic folds Cowork into Claude and adds Docs and Slides
Anthropic merged Claude Cowork and Claude chat into a single product on 16 September, removing the choice of which surface a task belongs in and letting Claude select the tooling itself. Claude Docs and Claude Slides launched alongside it in beta, and Claude Design now runs inside conversations; slides can be edited in place, presented from Claude, or downloaded as PowerPoint or PDF. It is rolling out to Pro and Max plans on web, desktop and mobile first, with Team and Free to follow. Existing Cowork chats, projects, artifacts, connectors and skills stay where they are.
Why it matters: Mode-switching was a tax users paid for a product boundary that was never theirs. Removing it is a bet that routing is a model decision, not a menu — and it puts document and deck generation in the default surface rather than an adjacent one.
Shanghai AI Lab puts a 744B agentic model under MIT
Atria Dawn Preview appeared on Hugging Face with an MIT licence and a technical report on arXiv dated 14 September. It is a 744B-parameter mixture-of-experts model built on GLM-5.2 with a 256K-token context, aimed at long-horizon work rather than single turns — problem analysis, tool use, code implementation, experiment execution, result analysis and failure recovery in one loop. The model card reports 92.5% on BrowseComp, 59.6% on SWE-bench Pro, 86.2% on MLE-bench Lite, 77.0% on BFCL v4 and 86.5% on CyberGym.
Why it matters: Permissively licensed weights aimed squarely at multi-day agent runs, with cybersecurity scores published rather than omitted. Verify the benchmarks yourself; the licence is the part that is not in dispute.
Cornelis raises $205M arguing half your GPU hours are spent waiting
Cornelis Networks announced $205 million in funding on 14 September alongside Active Compute Fabric, an open scale-up and scale-out network architecture combining lossless transport, in-fabric acceleration and programmable compute so the network can operate on data in flight and offload collective operations. The Intel Omni-Path spinout also announced a collaboration with Qualcomm on validation toward future rack-scale designs. Cornelis models a 100,000-GPU system as spending roughly 50% of GPU hours waiting for data — about $1.68 billion a year in idle capacity and 500 GWh of power.
Why it matters: The scarce resource in a large cluster is increasingly the interconnect, not the accelerator. If the 50% figure is even directionally right, networking is the cheapest capacity anyone can buy.
New on Onei this week
Released during this window and now in the catalogue.
Aside AppThe Windows build landed on Monday, under three months after the Mac launch — the interesting date, because the credential-vault argument only gets tested where the logged-in accounts actually live. Read it next to this week's Google Home item: two answers to letting an agent act as you.
Oats Appv0.25.0 shipped on Saturday and the changelog is all recording-durability work: audio checkpointed every five minutes, incremental transcript checkpoints, speaker-rename race conditions fixed. That is what a notetaker's issue tracker looks like once people trust it with meetings they cannot re-run.
mysetup.ai AppLaunched on Thursday and took 249 points and 138 comments on Hacker News the same day — by some distance the week's strongest reception. Read the comments before the site: the argument about whether published setups are useful signal or just tool fashion is the thing being tested here.
skillbay AppLaunched on Thursday as "Craigslist for agent skills, curated by a human". We run a skills directory, so treat this listing as interested: the question it raises is whether paid SKILL.md packages survive contact with a format that is a markdown file anyone can fork.
aclif AppAlso a Thursday launch. The bet worth judging is whether one command grammar across four very differently-shaped enterprise SaaS products stays coherent as the fifth and sixth arrive, or collapses into per-vendor special cases with a shared prefix.
MCPJam AppSix npm releases in four days, 3.6.0 through 3.8.1, which is the fastest cadence anything in this issue kept. If you maintain a server rather than consume one, the release velocity is the reason to look now rather than in a month.
Playwright CLI Skill0.1.20 on Monday and 0.1.21 on Friday — the only skill in the catalogue that cut a release inside the window, out of 18 we indexed this week. Everything else came from collections that predate it, which makes this a useful reference point for how rarely a skill actually ships.
- SAP Fiori MCP Server MCP server
1.12.6 and 1.12.7 landed midweek. A patch-level cadence on a vendor MCP server this deep into enterprise tooling is a better signal of real usage than any feature list — nobody ships two patches in two days for a demo.
Payload MCP Plugin MCP server3.90.0 and 3.90.1 shipped on Friday, with 4.0.0 canaries moving in parallel all week. Check which major line you are on before wiring an agent to it: the plugin tracks Payload's own version, so the MCP surface moves when the CMS does.