--- type: source-digest topic: ue-mcp-strategy status: active created: 2026-06-16 updated: 2026-06-16 source_notes: - external derived_from: - web-research confidence: medium source_url: https://www.figma.com/blog/introducing-figma-mcp-server/ --- # platform-ai-tooling-strategies ## Why This Matters Epic's build-vs-expose call is not novel — every adjacent platform / creative-tool company has already made a version of it (2025-2026). The cleanest signal is *which layer they protected*: the companies that opened an MCP did so where their money sits one layer below the authoring surface; the companies that built a closed harness did so where authoring IS the product or where the runtime is the consumption surface. This digest captures the public record so the claim in [[research/ue-mcp-strategy/concepts/harness-vs-mcp]] is auditable. ## Key Takeaways (one cluster per company, with URLs) **Figma — built BOTH, and the split is the tell.** - Shipped a *closed harness* at the authoring layer (Figma Make = prompt-to-code app builder; First Draft = prompt-to-design, becoming an agent entry point by May 2026). This is where Figma captures value, so it built and owns it. https://www.figma.com/blog/figma-ai-first-draft/ , https://www.figma.com/blog/design-systems-ai-mcp/ - Shipped an *open Dev Mode MCP server* (June 2025) at the design→code *handoff* — the layer Figma does NOT monetize (devs leave Figma to write code in Cursor/Copilot/Claude Code). Exposing it keeps frontier coding agents pointed at Figma files. https://www.figma.com/blog/introducing-figma-mcp-server/ - [NON-OBVIOUS] The MCP gets *more useful the more your design system is wired up* — i.e. the tool surface is engineered to emit Figma-advantaged output, not neutral output. https://www.figma.com/blog/design-systems-ai-mcp/ **Blender — community MCP, zero value capture at tooling (so no reason to close it).** - The viral "Blender MCP" was built by an outside individual, Siddharth Ahuja, not Blender. Open-source, dual-component (Blender addon socket server + MCP server). https://github.com/ahujasid/blender-mcp - Blender (a non-profit, GPL tool) captures no value at the authoring layer, so it had nothing to protect — the protocol got built *for* it by the community. Anthropic listed it as an official Claude connector on 2026-04-28 and donated to Blender's Python-API work; Blender took it as a one-time donation. https://www.pulsemcp.com/servers/ahujasid-blender , https://cline.bot/blog/human-intent-in-the-age-of-ai-insights-from-our-chat-with-siddharth-ahuja-creator-of-the-ableton-blender-mcps **Unity — the direct engine competitor: retreated from owning the model, opened an MCP.** - [NON-OBVIOUS] Unity *deprecated Muse* (its first-party models) and replaced it with "Unity AI," which uses *third-party frontier models* inside the editor — an explicit move from own-the-intelligence to rent-it. Shipped with Unity 6.2 (2025). https://iconpolls.com/blogs/unity-ai-review-2026-assistant-download-plugin-gdc-muse-user-experience-and-faqs , https://digitalproduction.com/2025/08/22/unity-6-2-welcomes-ai-but-pace-caution-user-liability-on-copyright/ - Built *official* first-party Unity MCP via its "AI Gateway" connecting Claude Code / Cursor to the editor, with managed permissions/entitlements. https://docs.unity3d.com/Packages/com.unity.ai.assistant@2.0/manual/unity-mcp-overview.html - Kept Sentis (renamed *Inference Engine*) — the *runtime* inference engine that ships models inside the shipped game. Note the split: opened the editor/authoring surface, kept the runtime piece. https://unity.com/features/ai - Thriving unofficial community MCPs predate the official one (CoderGamester/mcp-unity, CoplayDev/unity-mcp). https://github.com/CoderGamester/mcp-unity **Adobe — opened MCP everywhere its value is workflow/data, made it the default protocol.** - Connected Firefly/Photoshop/Express/Acrobat into ChatGPT *via MCP* (800M ChatGPT users), powered by Firefly Image Model 5 — Adobe still owns the *model*, exposes the *tools*. https://www.digitalapplied.com/blog/adobe-chatgpt-integration-creative-ai-guide - At Summit 2026 made MCP the sanctioned agent protocol for Adobe Commerce / Experience Cloud (rebranded "CX Enterprise"), shipped Commerce + Marketo MCP servers, on MCP + Google A2A. https://www.paz.ai/blog/adobe-mcp-commerce-default-agent-protocol , https://news.adobe.com/news/2026/04/adobe-redefines-custome-experience - AEM MCP servers publicly support ChatGPT, Claude, Cursor, Copilot Studio. https://experienceleague.adobe.com/en/docs/experience-manager-learn/cloud-service/ai/mcp-servers/overview - [NON-OBVIOUS] Adobe both *owns its model* (Firefly, "commercially safe") AND *exposes tools* via MCP — the two are not mutually exclusive when your model is itself a differentiated product. **GitHub Copilot — a built harness, because Microsoft owns the whole stack.** - Microsoft (owns GitHub) runs Copilot as an owned harness: in-editor assistant → autonomous coding agent (May 2025) → agent-native desktop app. https://github.com/newsroom/press-releases/coding-agent-for-github-copilot , https://github.blog/news-insights/product-news/github-copilot-app-the-agent-native-desktop-experience/ - It owns the *model relationship* (multi-model incl. OpenAI GPT-5.1) and ships a harness/SDK (Microsoft Agent Framework + Copilot SDK) — but it still *consumes* MCP servers as a host. https://www.microsoft.com/en-us/microsoft-365/blog/2025/11/18/microsoft-ignite-2025-copilot-and-agents-built-to-power-the-frontier-firm/ , https://devblogs.microsoft.com/agent-framework/build-ai-agents-with-github-copilot-sdk-and-microsoft-agent-framework/ - [NON-OBVIOUS] Copilot is a harness because GitHub's value capture IS the developer surface/subscription — the authoring layer is the product, so you must own it. **Roblox — in-house foundation model, because the runtime is the consumption surface.** - Built *Cube 3D*, its own foundation model for text→3D (and roadmap to "4D"), trained on *native 3D data* via a token-based autoregressive approach — own-the-intelligence at the generation layer. https://about.roblox.com/newsroom/2025/03/introducing-roblox-cube , https://techcrunch.com/2025/03/17/roblox-releases-its-open-source-model-that-can-create-3d-objects-using-ai/ - [NON-OBVIOUS] Roblox open-sourced the Cube *weights* (GitHub/HuggingFace) but the value capture is the *closed platform runtime* — generated assets monetize inside Roblox's engagement/economy, not via the model. Early 4D test ("Wish Master") saw +64% playtime, 160k+ objects generated — value lands at consumption, not authoring. https://about.roblox.com/newsroom/2026/02/accelerating-creation-powered-roblox-cube-foundation-model , https://github.com/Roblox/cube **Notion / Canva — opened hosted MCP; the workspace/runtime is the moat.** - Notion: open-source MCP (early 2025) → *hosted* MCP server tailored for agents (not raw REST), works with Claude/Cursor/ChatGPT; Custom Agents connect out to apps (Linear, Ramp, Canva) via MCP, gated to Business/Enterprise. https://www.notion.com/blog/notions-hosted-mcp-server-an-inside-look , https://www.notion.com/help/mcp-connections-for-custom-agents - [NON-OBVIOUS] Notion ships *LLM-friendly tool descriptions and "private" functionality slices* through the MCP — engineered-for-agents, not a thin API wrapper (same "advantaged surface" pattern as Figma). https://www.notion.com/blog/notions-hosted-mcp-server-an-inside-look - Canva appears as a *connected app* inside Notion's agent ecosystem, not a competitor — both bet on MCP as connective tissue while the workspace/data stays the moat. **The framing source — why the middle commoditizes (barbell).** - VC/enterprise analysis: MCP flattens the integration layer; lock-in shifts to model *choice* (now fungible). Value concentrates upstream (scale) and downstream (vertical/workflow), middle gets commoditized — the TCP/IP → Cisco+Netscape analogy. https://www.leoniscap.com/research/mcps-value-creation-capture-and-destruction-lessons-from-the-api-era - Anthropic's original MCP framing: an open standard to point frontier models at where data/tools live. https://www.anthropic.com/news/model-context-protocol ## Entities - Figma (Dev Mode MCP, Figma Make, First Draft) ; Blender + Siddharth Ahuja (blender-mcp) ; Unity (Muse→Unity AI, Sentis→Inference Engine, AI Gateway MCP) ; Adobe (Firefly, CX Enterprise, AEM/Commerce MCP) ; GitHub/Microsoft (Copilot, Agent Framework) ; Roblox (Cube 3D) ; Notion (hosted MCP) ; Canva ; Anthropic (MCP standard, Claude connectors) ; Leonis Capital (analysis). ## Sources (all URLs) - https://www.figma.com/blog/introducing-figma-mcp-server/ - https://www.figma.com/blog/figma-ai-first-draft/ - https://www.figma.com/blog/design-systems-ai-mcp/ - https://github.com/ahujasid/blender-mcp - https://www.pulsemcp.com/servers/ahujasid-blender - https://cline.bot/blog/human-intent-in-the-age-of-ai-insights-from-our-chat-with-siddharth-ahuja-creator-of-the-ableton-blender-mcps - https://docs.unity3d.com/Packages/com.unity.ai.assistant@2.0/manual/unity-mcp-overview.html - https://unity.com/features/ai - https://iconpolls.com/blogs/unity-ai-review-2026-assistant-download-plugin-gdc-muse-user-experience-and-faqs - https://digitalproduction.com/2025/08/22/unity-6-2-welcomes-ai-but-pace-caution-user-liability-on-copyright/ - https://github.com/CoderGamester/mcp-unity - https://www.digitalapplied.com/blog/adobe-chatgpt-integration-creative-ai-guide - https://www.paz.ai/blog/adobe-mcp-commerce-default-agent-protocol - https://news.adobe.com/news/2026/04/adobe-redefines-custome-experience - https://experienceleague.adobe.com/en/docs/experience-manager-learn/cloud-service/ai/mcp-servers/overview - https://github.com/newsroom/press-releases/coding-agent-for-github-copilot - https://github.blog/news-insights/product-news/github-copilot-app-the-agent-native-desktop-experience/ - https://www.microsoft.com/en-us/microsoft-365/blog/2025/11/18/microsoft-ignite-2025-copilot-and-agents-built-to-power-the-frontier-firm/ - https://devblogs.microsoft.com/agent-framework/build-ai-agents-with-github-copilot-sdk-and-microsoft-agent-framework/ - https://about.roblox.com/newsroom/2025/03/introducing-roblox-cube - https://techcrunch.com/2025/03/17/roblox-releases-its-open-source-model-that-can-create-3d-objects-using-ai/ - https://about.roblox.com/newsroom/2026/02/accelerating-creation-powered-roblox-cube-foundation-model - https://github.com/Roblox/cube - https://www.notion.com/blog/notions-hosted-mcp-server-an-inside-look - https://www.notion.com/help/mcp-connections-for-custom-agents - https://www.leoniscap.com/research/mcps-value-creation-capture-and-destruction-lessons-from-the-api-era - https://www.anthropic.com/news/model-context-protocol ## Open Questions - Did Unity drop Muse because owning a model was strategically wrong, or just because its first-party models lost to the frontier on quality/cost? (Motive matters — one is a strategy lesson, the other is a capability gap.) - Roblox owns a model AND opens weights AND runs a closed runtime — does that contradict the "pick one" framing, or confirm that value-capture-location (runtime) is what licenses every other choice? - Several framing claims (barbell, "model choice is now fungible") are VC commentary, not settled fact — flagged as medium confidence.