MCP Server
Get the best output
Practical rules for reliable agent changes.
- Read
get_guidelinesandget_editor_statebefore editing. Work on the focused design and ask for empty space when adding a top-level composition. Readget_design_systembefore making tokens or components andlist_commentsbefore revising reviewed work. - Read a shallow tree first, then exact IDs. Page wide documents with
offsetandlimit; page listings that traverse more than 1,000 nodes and responses over 256 KiB are refused. Usebatch_getwithdetail: "style"to inspect the editable values before replacing a fill, effect, layout, or constraint. - Batch operations for one visible component or screen section with a useful label. Content edits preview in order and commit as one undo step; a failed operation rolls content edits back. Selection and viewport changes are not transactional. Use returned created IDs in later batches.
- Use frames for clipping and auto-layout, groups for organization. Set container sizing and child participation explicitly where needed. Horizontal, vertical, and grid layouts are supported; use
set_grid_layoutfor columns and rows rather than hand-built nested rows when a true grid fits the content. - Screenshot the changed node or page, inspect it visually, and review the autosaved
.fnxdiff. - Edit
.fnxor managed project JSON directly for supported source changes outside the MCP operation vocabulary. Read all relevant source first, preserve node and component identities, validate the complete candidate withvalidate_fnx_source, and let it reload before touching the same nodes on canvas. - Import generated image files or URLs with
import_image, then place the saved asset using an appropriate editor or source workflow.prepare_design_assetonly opens a generation draft for user review. - Use
report_agent_activityfor visible milestones, including the actual page, node, coordinates, source path, and workspace. Reuse the same agent ID and name inbatch_design.activity; clear the activity withactive: falsewhen finished. Reply to comments in their threads and resolve only verified requests.
See the tool reference and alpha known issues.