Build1 distinct publisher3 min readUpdated
A bundled MCP server lets Claude Code, Cursor, Codex and Antigravity edit live scenes. There is no headless option, and no documented way back from exported code into the source file.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The undo history is doing more work here than the MCP server is. Spline says every change an agent makes is treated as a normal editor operation, landing in the undo stack and syncing to everyone else in the file [7]. That puts the rollback path in the same place a designer already looks, which is a cheaper form of review than inspecting whatever an agent wrote to disk.
Getting there requires the app to insert itself into your tooling. When the desktop application starts, it scans for supported clients and writes the server into their configuration files; the developer then restarts the client and prompts normally [4]. Five clients are covered [15], and a single connection spans two surfaces, with Spline deciding whether a request is scene work for the 3D editor or interface work for Hana [12]. Convenient, and also a reminder that discovery is the vendor's call rather than the client's.
The asymmetry is on the way out. Spline's runtime API can control a published scene, but the company documents no path for changes made in exported code to flow back into the original 3D file; an agent that needs to revise the source has to return through MCP and work in the live editor [9]. Set that against the packaging and you are left with one writable entrance to the source of truth, which is an MCP call executed against an open tab on somebody's desktop [16]. Anything an agent edits after export is a branch with no merge.
Hana's half of the round trip is narrower than it sounds. The agent can turn an existing component into an editable frame and carry visual changes back into application code, using stable references on exported elements to find the right place in the source [10]. Hana supports only a subset of HTML and CSS built around flexbox, though, so complex markup and framework-specific components have to be reconciled with the existing code rather than replaced outright [11]. Reconciliation is human work, and it is where the closed loop between design file and repository actually leaks.
Spline already ships its own agent that will build and edit a scene from a prompt [14], so the pitch for MCP is not that a model can produce 3D. It is that the coding agent already holds the codebase and the work in progress, and nobody wants to rebuild that context in a separate Spline conversation [c2b]. That is an argument about where context should live, and it has a consequence the marketing does not mention: the context stays in the client, while the artefact stays in Spline, and only Spline's application can carry a write between them.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Spline says its internal AI agent follows the same approach, with each change treated as a normal editor operation that appears in the undo history and syncs to everyone else working in the file.
Spline's runtime API can control a published scene, but the company does not document a way for changes made to the exported code to flow back into the original 3D file; to revise the source, the agent has to return through MCP and work in the live editor.
Spline V2 also includes its own AI agent, which can build and edit a scene from a prompt.
Spline released V2 on Thursday, a complete rebuild of its 3D editor that enables external coding agents to work directly on live, editable scenes.
The Spline MCP Server opens the editor to Claude Code, Cursor, Codex, Google Antigravity and VS Code, which can change a live Spline project's scene and behaviour while keeping the result fully editable.
MCP lets developers access the same design environment from a coding agent that already understands the codebase and the work in progress, without rebuilding that context in a separate Spline conversation.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single trade report on vendor documentation
Every claim traces to one article from one publisher, and its substance is drawn from Spline's release material and documentation rather than independent testing. The reporting is unusually specific about mechanics and explicit about what the documentation does not cover, which raises credibility, but nothing here is corroborated by a second outlet, a hands-on trial, or a measurement.
Launch announcements only
All adoption signal in the cluster is release-shaped: the V2 launch, the desktop-only bundling of the MCP server, and adjacent releases from ElevenLabs, Antigravity and Codex. There are no install counts, customer names, deployment reports, or usage disclosures, and desktop-only distribution structurally excludes automated or server-side use, so the measured value stays near the floor.
Hot headline, caveated body
The framing that Spline 'handed the keys to Claude Code' overstates a capability that is documented but untested, and the built-in-agent parity claim is carried on the vendor's word. The body largely closes the gap by naming the ceilings itself — desktop-only, no headless server, no return path from exported code, Hana's flexbox-only subset, unaddressed change attribution — so the overstatement is mild rather than structural.
Vendor-timed launch coverage
The article is launch-cycle coverage of a vendor's product event, with key assertions sourced to Spline ('Spline says', 'According to Spline') and no independent verification. Spline benefits directly from positioning its editor as the surface agents write into. Offsetting this, the publisher volunteers unfavorable findings — the missing export-to-source path and the undocumented review/attribution flow — and no sponsorship or commercial relationship is disclosed in the supplied material.
Coherent but single-sourced
The mechanical description is internally consistent, specific, and self-limiting, which supports moderate confidence in the architectural facts — especially the desktop-only door and the absent write-back path. Confidence is capped by the single publisher, the reliance on vendor documentation for behavioral guarantees, and the total absence of adoption measurement.
build
Your Coding Agent Reads .env, And .gitignore Was Never The Control1 distinct publisher
build
Claude Code now outruns Copilot roughly two to one in JetBrains' survey of 15,000 developers1 distinct publisher
science
OX Security says MCP command execution is a design choice, so server owners own the risk1 distinct publisher
invest
Binance gives AI agents their own subaccounts, and no loss limit1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 21, 2026