tldraw-mcp
Minimal MCP server for editing tldraw .tldr files via JSON manipulation. Headless, no browser needed.
README
tldraw-mcp
Minimal MCP server for editing tldraw .tldr files via JSON manipulation. Headless, no browser needed.
Status
Working skeleton. Schema validation is wired (@tldraw/tlschema validators run before every write), fractional indexing uses @tldraw/utils, file writes are guarded by proper-lockfile. Output verified end-to-end against the real tldraw runtime via Store.loadStoreSnapshot() in the contract test layer.
Quick Start
Requires Node ≥ 20.
1. Install (user scope — available in every project):
claude mcp remove tldraw-m9810223 -s user 2>/dev/null; rm -rf ~/.npm/_npx
claude mcp add -s user tldraw-m9810223 -- npx -y github:m9810223/tldraw-mcp
Restart Claude Code. /mcp should now list tldraw-m9810223.
2. Try the demo prompt in Claude Code:
Draw the Shai-Hulud supply-chain attack flow described in this article with MCP `tldraw-m9810223`, save to `./demo.tldr`:
https://semgrep.dev/blog/2026/malicious-dependency-in-pytorch-lightning-used-for-ai-training/
- Match the actors and how data/control flows between them
- Polish the result at the end
3. View the result — drop ./demo.tldr onto tldraw.com, or use the tldraw VS Code extension for live preview.

Source file: docs/demo.tldr
Tools
File / page lifecycle
| Tool | What it does |
|---|---|
create_empty_file |
Create a fresh .tldr with a default page |
create_page |
Add a new page |
list_pages |
List pages with id, name, ordering index |
move_to_page |
Move shapes; bindings: 'error' | 'pull' | 'cut' controls binding handling |
Shapes
| Tool | What it does |
|---|---|
create_rect |
Create a rectangle (geo shape) |
create_text |
Create a text shape |
create_group |
Group shapes by reparenting them |
ungroup |
Dissolve a group, reparenting its children to the group's parent |
connect |
Arrow + bindings between two same-page shapes |
list_shapes |
List shapes — id, type, x, y, label only |
get_shape |
Full record of one shape by id |
update_shape |
Shallow-merge patch (use nested { "props": {...} } for prop edits) |
delete_shape |
Delete by id; cascade: true (default) also removes attached arrows + bindings |
Layout / text fitting
| Tool | What it does |
|---|---|
fit_to_text |
Resize a geo/text shape to fit its current text content |
align |
Align shapes on an axis (left/right/top/bottom/center-x/center-y) |
distribute |
Even-space shapes between the outermost two |
auto_layout |
Lay shapes out in a horizontal/vertical chain |
graph_layout |
Dagre layout for arrow-connected shapes (best for non-chain topologies) |
measure_arrow_labels |
Report label sizes + endpoint distances for labeled arrows |
bend_overlapping_arrows |
Bend parallel arrows (same shape pair) symmetrically; priority[] keeps important arrows straight. graph_layout calls this automatically. |
polish_layout |
One-shot finisher: fit_to_text every node + graph_layout (auto-bends arrows). Use as the last step after building a fresh diagram. |
Discovery & escape hatch
| Tool | What it does | Token cost |
|---|---|---|
search_api |
List supported shape types + curated required props. Pass {type, verbose:true} to dump live prop names from @tldraw/tlschema for any type (including ones not in the curated list) |
low / medium |
exec_jq |
Run a jq filter against the file. write=true persists (auto-checkpoint first) |
varies |
Checkpoints (safety)
| Tool | What it does | Token cost |
|---|---|---|
save_checkpoint |
Copy .tldr to a timestamped backup |
low |
list_checkpoints |
List backups, newest first | low |
restore_checkpoint |
Restore a backup (most recent if checkpoint omitted) |
low |
The token-saving design: tools take primitive args, return ids or ok. The full JSON only enters context when you call get_shape deliberately.
Install / Update / Remove
jq only needed for exec_jq (brew install jq / apt-get install jq).
# Install — user scope: every project on this machine (recommended)
claude mcp remove tldraw-m9810223 -s user 2>/dev/null; rm -rf ~/.npm/_npx
claude mcp add -s user tldraw-m9810223 -- npx -y github:m9810223/tldraw-mcp
# Install — local scope (default): only the current project directory
claude mcp remove tldraw-m9810223 2>/dev/null; rm -rf ~/.npm/_npx
claude mcp add tldraw-m9810223 -- npx -y github:m9810223/tldraw-mcp
The first arg (tldraw-m9810223) is the local server name — pick whatever you like, then refer to it the same way in subsequent commands. Restart Claude Code, then /mcp lists it with 26 tools.
Wire up to other MCP clients
Same JSON shape, different config file location:
| Client | Config path |
|---|---|
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) |
| Cursor | ~/.cursor/mcp.json |
| VS Code | .vscode/mcp.json |
Bootstrapping a .tldr file
Use the create_empty_file tool, or save an empty canvas from tldraw.com and point tools at the absolute path.
create_empty_file({ file: "/tmp/demo.tldr" })
Viewing / editing the output
| Tool | Notes |
|---|---|
| tldraw.com | Drag the .tldr file onto the page |
| tldraw VS Code extension | Native .tldr preview + edit inside VS Code; survives file changes from the MCP server |
| Tldraw Desktop / official editor | Drop the file in |
The VS Code extension is the smoothest dev loop — keep code path/to/file.tldr open in a tab while the MCP edits it; the editor refreshes on disk change.
Design comparison vs official tldraw-mcp-app
The Cloudflare-hosted official MCP exposes only search + exec (run any JS in a live tldraw Editor). This skeleton goes the opposite way — typed JSON edits over .tldr files — and borrows the discovery pattern (search_api) and escape hatch (exec_jq) so an LLM can fall through when typed tools don't cover an operation.
Official tldraw-mcp-app |
This skeleton | |
|---|---|---|
| Transport | streamable-http + sse (Cloudflare) | stdio (works in Claude Code directly) |
| Runtime | Real tldraw Editor in widget iframe | Pure Node, edits raw JSON |
| Tools | 2 (search, exec) + checkpoints |
26: file/page (4) + shape (9) + layout (8) + discovery (2) + ckpt (3) |
| Live preview | Yes (widget iframe) | No (open the file in tldraw to view) |
| Coverage | Whole Editor API | Geo / text / arrow + jq escape hatch |
Known gaps
index(z-order) only supports appending above the current max — no insert-between- No image / video / asset support
- Schema version pinning is informational only — opening a file in a newer tldraw may trigger migrations
search_apicurated list is hand-maintained alongside the live@tldraw/tlschemareflection
Architecture
src/
index.ts MCP server entry, tool registration (stdio transport)
tools.ts Tool handlers + zod input schemas
shapes.ts tldraw record factories (geo/text/arrow/group/binding)
store.ts Load/save .tldr + withFileLock; helpers (id gen, indexing, find, page-of-shape, bindings-for-shape)
template.ts Empty .tldr generator using @tldraw/tlschema serialize()
validate.ts validateShape / validateBinding using createShapeValidator + createBindingValidator
checkpoint.ts Timestamped backups under .tldraw-mcp-checkpoints/
jq.ts Shell-out to jq for the exec_jq escape hatch
text-metrics.ts Text size heuristics for autoFit + label measurement
graph-layout.ts Dagre wrapper for graph_layout
arrow-bending.ts Detect parallel-edge overlap + assign symmetric bend values
test/
unit/ store + validate + text-metrics + arrow-bending (34 tests)
integration/ tools end-to-end on tmp .tldr (39 tests)
contract/ loadStoreSnapshot against real @tldraw/store (4 tests)
boundary/ N-1 / N / N+1 limits + concurrent writes (89 tests)
Pure JSON manipulation — no @tldraw/store, no DOM, no React.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。