`/models` now returns only deepseek-v4-pro and deepseek-v4-flash;
deepseek-chat / deepseek-reasoner sunset 2026-07-24 and the
deepseek-v3.2 hard-coded in the ark-coding fallback list never
existed. Both v4 models default to thinking enabled and the API
toggles via `{"thinking":{"type":"disabled"}}` — keep
`thinkingMode: 'disabled'` so the app's fast/non-thinking default
stays intact (server reasoning paths honor it; the Zig openai-compat
path doesn't emit the toggle yet, so calls through that path still
get provider-default thinking until it's wired). v4-pro promoted to
full tier; legacy aliases pinned to an exact RegExp so future
deepseek-* variants don't inherit a forced disabled mode.
Bandaid for the unwired toggle: v4-pro gets `timeoutMultiplier: 2`
because its default-on reasoning blows past the orchestrator's
planning timeout on long system prompts (observed in dev: planning
phase falls back, sub-agent then succeeds — UX degraded but
functional). Drop the multiplier once the Zig path actually sends
`thinking:{type:disabled}`.
Don't add a BUILTIN_MODEL_LISTS.deepseek entry — DeepSeek exposes
/v1/models, so let `fetchProviderModels` pull the live catalog
through `/api/ai/provider-models` instead of pinning a snapshot
(the ark-coding `deepseek-v3.2` ghost above shows what those
snapshots drift into).
|
||
|---|---|---|
| .. | ||
| src | ||
| CLAUDE.md | ||
| LICENSE | ||
| package.json | ||
| README.md | ||
| tsconfig.json | ||
@zseven-w/pen-mcp
MCP server for OpenPencil — enables Claude, GPT, Gemini, and other LLMs to read, create, and modify designs through a standard tool protocol.
Install
npm install @zseven-w/pen-mcp
# or
bun add @zseven-w/pen-mcp
Overview
pen-mcp exposes OpenPencil's full editing API as MCP tools. External AI agents can open documents, inspect the canvas, insert/update/delete nodes, and generate complete designs — all through structured tool calls.
Three workflows are supported:
| Workflow | Tools | Best for |
|---|---|---|
| Single-shot | insert_node, batch_design |
Quick edits, single components |
| Layered | design_skeleton → design_content × N → design_refine |
Full-page designs with high fidelity |
| CRUD | batch_get → update_node / delete_node |
Reading & modifying existing content |
Quick Start
# Run as stdio MCP server (for Claude Desktop, Cursor, etc.)
npx @zseven-w/pen-mcp
# Or connect to a running OpenPencil instance
op mcp:dev
Claude Desktop Configuration
{
"mcpServers": {
"openpencil": {
"command": "npx",
"args": ["@zseven-w/pen-mcp"]
}
}
}
Tools
Document & Read Tools
| Tool | Description |
|---|---|
open_document |
Open an .op file or connect to the live Electron canvas. Always call first. |
batch_get |
Search and read nodes by type, name regex, or specific IDs. Controls read depth for nested content. |
get_selection |
Get the currently selected nodes on the live canvas. |
snapshot_layout |
Get a compact bounding-box layout tree — useful for spatial understanding. |
find_empty_space |
Find available canvas space in a given direction for placing new content. |
get_design_prompt |
Retrieve segmented design knowledge (schema, layout, roles, text, style, icons, examples). |
Node CRUD Tools
| Tool | Description |
|---|---|
insert_node |
Insert a new node with full PenNode data. Supports postProcess for auto-defaults. |
update_node |
Shallow-merge properties into an existing node. |
delete_node |
Delete a node and all its children. |
move_node |
Reparent a node to a new container. |
copy_node |
Deep-clone a node with new IDs under a target parent. |
replace_node |
Replace a node entirely with new data at the same position. |
import_svg |
Import a local SVG file as editable PenNodes. |
Batch Design DSL
batch_design accepts a compact DSL — one operation per line:
root=I(null, { "type": "frame", "name": "Page", "width": 1200, ... })
header=I(root, { "type": "frame", "name": "Header", ... })
U(header, { "fill": [{ "type": "solid", "color": "#1A1A2E" }] })
logo=C("existing-logo", header, { "x": 24 })
M("floating-btn", header)
D("old-section")
| Op | Syntax | Description |
|---|---|---|
| I | binding=I(parent, { data }) |
Insert node |
| U | U(path, { updates }) |
Update properties |
| C | binding=C(source, parent, { overrides }) |
Copy node |
| R | binding=R(path, { newData }) |
Replace node |
| M | M(nodeId, parent, index?) |
Move node |
| D | D(nodeId) |
Delete node |
Layered Generation Workflow
For high-fidelity multi-section designs:
1. design_skeleton → Create root frame + section placeholders
2. design_content → Fill each section with content nodes (call per section)
3. design_refine → Run full-tree validation and auto-fixes
Page Management
| Tool | Description |
|---|---|
add_page |
Add a new page to the document |
remove_page |
Remove a page (cannot remove the last one) |
rename_page |
Rename a page |
reorder_page |
Move a page to a new index |
duplicate_page |
Deep-clone a page with new IDs |
Post-Processing
All creation tools support postProcess=true for automatic:
- Semantic role defaults (button padding, input height, card radius, etc.)
- Icon name → SVG path resolution (Lucide icon set)
- Card row equalization in horizontal layouts
- Text height estimation
- Frame height expansion when content overflows
clipContentauto-addition for frames withcornerRadius+ images
Design Prompt Sections
get_design_prompt(section) returns focused subsets of design knowledge:
| Section | Content |
|---|---|
schema |
PenNode type definitions and property reference |
layout |
Flexbox layout engine rules (gap, padding, justify, align) |
roles |
Semantic roles and their auto-defaults (button, input, card, navbar, ...) |
text |
Typography rules, CJK support, copywriting guidelines |
style |
Visual style policy (colors, fonts, aesthetic) |
icons |
Feather/Lucide icon naming conventions |
examples |
Complete design examples with DSL |
guidelines |
Design tips (cards, inputs, phone mockups, hero sections) |
planning |
Layered workflow guide with section decomposition rules |
Live Canvas Sync
When connected to a running OpenPencil desktop app, changes made via MCP tools appear on the canvas in real-time. The sync is bidirectional — user edits on the canvas are reflected in subsequent batch_get / snapshot_layout calls.
Programmatic Usage
import { configureMcpHooks, MCP_DEFAULT_PORT } from '@zseven-w/pen-mcp';
// Configure custom hooks (optional)
configureMcpHooks({
onDocumentOpen: (path) => console.log(`Opened: ${path}`),
onNodeInsert: (node) => console.log(`Inserted: ${node.id}`),
});