--- type: concept topic: ue-mcp-strategy status: active created: 2026-06-16 updated: 2026-06-16 source_notes: - "[[research/ue-mcp-strategy/sources/platform-ai-tooling-strategies]]" derived_from: - web-research confidence: medium shared_concept: --- # Harness vs MCP (own the intelligence vs rent the frontier's) ## The Core Distinction (plain terms) A **harness** = you build and own the agent loop, the model relationship, and the authoring UX. You are competing on *intelligence* — a layer that improves every few months and that frontier labs (OpenAI, Anthropic, Google) dominate by spending billions. Win here only if you out-build the frontier or have proprietary signal it lacks. An **MCP** = you expose a *tool surface* and let frontier models drive. You rent their intelligence "for free," ride their improvement curve, and keep them pointed at *your* platform. You give up control of the loop in exchange for never having to win the model race. The decision rule is one sentence: **own the intelligence only if you capture value at the intelligence/authoring layer; otherwise expose tools and let the authoring layer commoditize.** ## Why It Matters for Epic Epic's money is one layer *below* authoring — the UE 5% runtime royalty, Fortnite/UEFN engagement, Fab. If that holds (see [[research/ue-mcp-strategy/concepts/ue-value-capture-map]]), then authoring is *free to commoditize* and a harness is a tax: Epic would spend to win a model race it does not need to win, when an MCP lets the frontier improve UE authoring at zero marginal cost. The one constraint: the MCP must emit **UE-advantaged** content that deepens [[research/ue-mcp-strategy/concepts/runtime-gravity-and-lock-in]], never engine-neutral content (see [[research/ue-mcp-strategy/concepts/mcp-as-tooling-surface]]). ## What Comparable Companies Chose (and why) - **Figma** built a *harness at authoring* (Figma Make, First Draft — that IS the product) but opened an *MCP at the design→code handoff* it does not monetize. Split by value layer. - **Unity**, the direct engine rival, *killed its own model* (Muse) and switched to third-party frontier models + an official MCP — the loudest own→rent signal in the set — while keeping Sentis/Inference Engine at the *runtime*. - **Roblox** built an in-house model (Cube 3D) because its *closed runtime is the consumption surface* — generation monetizes through engagement, so owning the generator is owning value. - **GitHub Copilot** is a harness because Microsoft owns the developer surface and the model relationship — authoring is the product. - **Adobe** owns its model (Firefly) *and* exposes MCP everywhere, because a differentiated model is itself a sellable product. - **Blender** (no value capture at tooling) didn't build anything — the community shipped its MCP *for* it. ## The Pattern [NON-OBVIOUS] You build a **harness** when authoring/runtime IS your value capture (Roblox runtime, Copilot/Microsoft surface, Figma's design app). You open an **MCP** when your value sits a layer *below* the authoring surface and the frontier can improve that surface for you (Figma's handoff, Unity's editor, Adobe's data/workflow, Notion's workspace). The map is not "creative tool vs not" — it is **where the dollar lands relative to the authoring layer.** Epic's dollar lands *below* authoring, which places it with the MCP-openers, not the harness-builders. ## Tensions / Contradictions - **Roblox owns a model, opens its weights, AND runs a closed runtime** — it does not "pick one." That either breaks the framing or proves the deeper rule: owning the *runtime* licenses every other choice. - An open MCP invites an **engine-agnostic standard** that could commoditize UE into a swappable backend — exactly the risk in [[research/ue-mcp-strategy/concepts/the-bear-case]]. Renting the frontier and being commoditized by it are the same move viewed from two sides. - A thin harness may still be needed for the *median* creator even if MCP is the strategic bet — either/or may be a false frame. ## Critical Questions (where this could be wrong) - Did Unity drop Muse on *strategy* or because its model simply lost on quality/cost? If the latter, Epic's "don't own intelligence" lesson is weaker than it looks. - Does Epic actually capture the AI-content wave at the *runtime royalty*, or does that value flow through Fortnite/UEFN engagement — a different team's decision? (Unresolved; see value-capture map.) - "Exposing tools makes frontier models better at UE authoring" is an attractive flywheel that may be more hope than mechanism (see [[research/ue-mcp-strategy/concepts/ai-content-generation-frontier]]). ## Related Concepts [[research/ue-mcp-strategy/concepts/ue-value-capture-map]] · [[research/ue-mcp-strategy/concepts/mcp-as-tooling-surface]] · [[research/ue-mcp-strategy/concepts/runtime-gravity-and-lock-in]] · [[research/ue-mcp-strategy/concepts/ai-content-generation-frontier]] · [[research/ue-mcp-strategy/concepts/ue-runtime-architecture]] · [[research/ue-mcp-strategy/concepts/the-bear-case]] ## Sources [[research/ue-mcp-strategy/sources/platform-ai-tooling-strategies]] — full URL list there. Primary anchors: Figma Dev Mode MCP (figma.com/blog/introducing-figma-mcp-server), Unity Muse→Unity AI (iconpolls / digitalproduction), Roblox Cube (about.roblox.com/newsroom/2025/03), Leonis Capital barbell analysis (leoniscap.com), Anthropic MCP (anthropic.com/news/model-context-protocol).