glyphs-mcp-scripting

Glyphs MCP scripting

Turn a natural-language idea into the smallest verifiable Python script for the running Glyphs app.

Core rules

• Prefer an existing dedicated MCP tool or domain skill when it already covers the task. Keep outline-specific fallback code in glyphs-mcp-outlines-docs and guarded italic fallback code in glyphs-mcp-italic-first-pass.
• Read the live server, font, master, glyph, layer, and selection context needed for the request before writing code.
• Search with docs_search, then fetch only the relevant pages with docs_get before using an unfamiliar GlyphsApp API. Target Glyphs 3.5 and Glyphs 4 unless the user requests one version.
• Use execute_code_with_context for glyph- or layer-scoped work. Use execute_code only when injected font, glyph, and layer context is not useful.
• Keep code focused, validate every target, make repeat runs safe when practical, and bound returned output with max_output_chars and max_error_chars.
• Never call exit(), quit(), or sys.exit(). Do not install or reload scripts or plug-ins, restart Glyphs, save a font, access files or the network, or launch subprocesses without separate explicit authorization.
• Never treat printed output as mutation proof. Re-read affected state with dedicated MCP tools and optionally inspect get_document_change_overview.

Workflow

1. Form a compact brief from the request:
   • exact target and scope
   • expected behavior
   • permitted effects
   • observable success condition
2. Resolve missing live context with get_server_info, list_open_fonts, get_selected_font_and_master, and the smallest relevant selection or inspection tools. Ask only when a consequential target or scope remains ambiguous.
3. Search and fetch the focused API documentation needed for the script.
4. Write the smallest working vertical slice. Validate object existence before reading or changing it, and summarize outcomes as changed, skipped, and failed counts when processing a batch.
5. Classify the exact code before execution:
   • For clearly read-only code that the user asked to run, execute it and inspect the bounded result.
   • For any font mutation or external side effect, first call the matching execution tool with snippet_only=true. Show the exact code, targets, and expected changes, then stop for explicit approval.
6. Treat the submitted code argument as the reviewed executable. Label the returned Macro Panel snippet separately because it wraps that code for manual use. Bind approval to the exact execution tool, code, font/glyph context, remaining arguments, targets, and side-effect scope.
7. After approval, execute only that unchanged reviewed request. If any bound field changes, generate a new snippet and request approval again.
8. For larger layer edits, pair layer.beginChanges() and layer.endChanges() in try/finally. Avoid MCP-driven glyph.beginUndo()/glyph.endUndo() because live Glyphs 4 testing found that those groups can trigger the undo recovery dialog.
9. Re-read the affected state, compare it with the brief's success condition, and report what ran, what changed, what was skipped, errors, and whether the font remains unsaved.
10. On a traceback, reduce to a minimal reproducer and retry once. Stop after a repeated failure and report any state that may already have changed.