AlphaSignal — Claude creative MCPs, MIT Platonic Hypothesis, Copilot token billing — @Lior Alexander (2026-04-28)

Why this is in the vault

Headline (Claude + 9 creative-tool connectors via MCP) is the announcement RDCO is already tracking for the Squarely book-cover-via-Adobe pilot. Two other items are load-bearing: MIT’s Platonic Representation Hypothesis (multi-model convergence on a shared internal world map — affects how we think about model substitutability and the long-run moat of any single-vendor lock-in) and GitHub Copilot’s switch to per-token billing on June 1st (validation of the agentic-session-cost thesis Lior keeps repeating; the per-seat-SaaS-billing model breaks when usage scales with autonomy).

Sponsorship

Three third-party paid slots, all explicitly disclosed:

• WorkOS — “How SSO works and the fastest way to add it.” Standard infra advertorial. Third-party paid, no editorial blending.
• Agent Field AI — pitching a “harness primitive” that composes harness systems into autonomous software. SWE-AF (100+ harness software factory) and cloud-security agents. Apache 2.0. Third-party paid. Notable for the terminology overlap with RDCO’s harness-thesis cluster — this is somebody else trying to commercialize “harness” as a primitive.
• AugmentCode — “AI coding tools aren’t enough — you need proof. Get the blueprint to turn AI into measurable ROI.” Lives inside the Signals list as item #2. Third-party paid, classic ROI-anxiety play.

No sponsor disclosure is hidden. All clearly marked “Presented by” / “partner with us”.

Issue contents

Top News — Claude connects to Blender, Adobe CC, Canva, Ableton, +5 more

9 new MCP-based connectors for creative tools. The framing: stay in Claude, talk to the tool directly, no app-switching. Specific connectors called out:

• Adobe Creative Cloud — multi-step workflows across Photoshop, Illustrator, Firefly, Premiere
• Blender — natural-language wrapper around the Python API; can analyze/debug entire scenes, batch-modify objects, register new tools in the Blender UI
• Autodesk Fusion — conversational 3D modeling
• Affinity by Canva — batch image adjustments, layer renaming, file export
• Splice — search royalty-free sample catalog from inside Claude
• Ableton, SketchUp, Resolume also included

Built on MCP, so accessible to other models too — this is Anthropic seeding the MCP ecosystem on the creative-tool side after the dev-tool side already saturated.

Top Repo — OpenAI realtime-voice-component

Apache-2.0 React package on gpt-realtime-1.5. Voice control for app actions: forms filled by speaking, settings switchers, multi-step flows where voice replaces tapping. You define narrow actions, voice layer calls them, your existing state logic handles the rest.

Top Repo — Gemma 4 local browser agent

Fully in-browser AI agent on Gemma 4 E2B running via WebGPU + Transformers.js. No cloud, no API keys, no server. Plain-English browsing-history search, summarize current page, tab management. Models cached after first download. Privacy story: tabs, history, page content never leave device.

Signals (6 numbered items)

1. Microsoft free tool: single photo to 3D model in seconds
2. (Sponsor) AugmentCode AI-ROI blueprint
3. CorridorKey neural-net green-screen keying — perfect keys per pixel (11,833 stars)
4. MIT Platonic Representation Hypothesis — major models converge on the same internal reality map regardless of training data (2,309 likes)
5. Qwen3 35B MoE distilled from Claude Opus shipped as free quantized GGUF (9,382 downloads)
6. GitHub Copilot moves to pay-per-use billing starting June 1st — agentic sessions eat compute, seat pricing breaks (2,783 likes)

Mapping against Ray Data Co

Strong, multiple vectors:

1. Claude creative MCPs → Squarely pilot is already live conversation. Today’s RDCO discussion was the book-cover-via-Adobe pilot using exactly this connector. Direct validation of the timing — we’re moving the day Anthropic ships, not waiting six months for tooling to mature. Affinity by Canva connector also matters: most KDP cover work touches Canva-tier tools, and “batch image adjustments + layer renaming + file export” is exactly the Squarely book-cover production loop. Worth a focused experiment within the next two weeks while the connector is fresh.

2. MCP-as-protocol is winning. “Built on MCP, accessible to other models too” — Anthropic is explicitly playing the protocol game, not the lock-in game. Reinforces the agent-deployer thesis: betting on the connector layer (MCP) over any specific model (Claude) is the durable position. This is the third issue in two weeks where MCP is the load-bearing primitive, not Claude itself.

3. MIT Platonic Hypothesis = quiet existential threat to “which model do I pick” framing. If different models trained on different data converge on the same internal map, the model-tier-as-financial-variable argument from yesterday’s 2026-04-27-alphasignal-anthropic-claude-marketplace-agent-quality needs nuance: the convergence claim says capability gaps narrow over time, but the empirical Anthropic marketplace data says today’s gap is real and costly. Both can be true. RDCO position: don’t bet the COO architecture on Claude-specific behavior; bet on the protocol (MCP) and the harness, swap models as economics change. This fits the Sanity Check editorial angle — “the AIs aren’t different, but the bills are.”

4. Copilot June-1 token billing = full validation of Lior’s intro framing. “Agentic sessions eat more compute” → seat-based SaaS pricing collapses. This matters for any RDCO product that sells access to a Ray-style agent: the per-seat model is dead-on-arrival for autonomous use cases. Implication for MAC (info-product) and any future Squarely-agent surface: usage-based billing from day one, not seat pricing retrofitted later. Worth a Sanity Check angle: “what your AI bill will look like in 18 months.”

5. Agent Field AI sponsor = competitive signal in the harness space. Someone is trying to sell “harness primitive” as a productized concept. Their pitch (SWE-AF: 100+ harness software factory) is closer to Devin-style code generation than to RDCO’s COO-as-agent thesis, but the vocabulary collision matters. If “harness” becomes a category term the market recognizes, RDCO’s harness-thesis cluster (2026-04-15-thariq-claude-code-session-management-1m-context and downstream) gets either lift or noise. Watch.

Medium:

6. Gemma 4 local browser agent reinforces the on-device + privacy thread already in the vault from the 2026-04-19-alphasignal-gemma-4-orchestration note. Not changing anything today; useful awareness for any future “private founder agent” surface.

7. OpenAI realtime-voice-component — not on RDCO’s stack, but the pattern (voice → narrow defined actions → existing state logic) is the right shape if we ever want voice control on the COO interface. File for awareness.

Skip:

• CorridorKey green-screen neural net (no RDCO surface touches video keying).
• Microsoft photo-to-3D (cute, no near-term application).
• Qwen3 35B distilled-from-Opus (interesting model availability fact; not running local 35B MoE today).

Curation section — notes

• All headline + Top Repo links go to third-party domains (Anthropic, OpenAI repo, Hugging Face / Chrome Store). No self-cross-promo from AlphaSignal itself.
• Three sponsor blocks (WorkOS, Agent Field AI, AugmentCode) all explicit and labeled.
• AugmentCode placement inside the Signals list as item #2 is the kind of editorial-blending pattern worth flagging — readers scanning the numbered list could miss the “Presented by” line. Not deceptive, but worth noting for our own newsletter design.
• No tracked-author candidates surfaced this issue.

Deep-fetch decisions

Cap is 2. Skipping all deep-fetches:

• Anthropic creative MCP announcement — RDCO already has the parallel conversation context on this; the AlphaSignal blurb plus Anthropic’s own settings-page is what we need. Founder can pull the official Anthropic post on demand if a Sanity Check angle materializes.
• MIT Platonic Hypothesis — would deep-fetch the arXiv paper if we were writing about it directly. Today the framing is sufficient. Queue as a candidate research question if the COO-architecture argument needs the citation.
• Copilot billing change — the fact alone (June 1, per-token) is the load-bearing data point. Microsoft’s announcement post is one click away if needed.

Related

• 2026-04-27-alphasignal-anthropic-claude-marketplace-agent-quality — yesterday’s issue; model-tier-as-financial-variable argument that the Platonic Hypothesis nuances
• 2026-04-19-alphasignal-gemma-4-orchestration — prior Gemma 4 framing
• 2026-04-15-alphasignal-anthropic-routines-claude-code — adjacent Anthropic routines / Claude Code coverage
• 2026-04-09-alphasignal-meta-muse-spark-anthropic-managed-agents — Anthropic Managed Agents
• 2026-04-25-squarely-current-state-review — Squarely surface that the Adobe/Affinity connector touches directly
• 2026-04-04-anthropic-skills-internally — Anthropic operating norms
• 2026-04-26-alphasignal-deepseek-v4-kimi-k26-agentic-ai — adjacent agentic-AI roundup

Copyright note

Paraphrased throughout. Single short quotes ≤15 words: “we’re not building a zoo of AI tools,” “the tools are merging, the costs are scaling.”