Run your Miro workshops, retros, and planning sessions from the command line. Read boards, create stickies, organize frames, and generate diagrams through Claude, Cursor, or any MCP client.
Community project — Not officially affiliated with Miro. See official options below.
MCP server for controlling Miro whiteboards with AI assistants
stdio
授權條款
MIT
API 金鑰
MIRO_ACCESS_TOKEN
上次檢查
2026-11-17
星數
33
工具
—
可用率
—
110 tools | Single binary | All platforms | All major AI tools
For Claude clients that support skills (Claude Code, Claude Desktop), this repo includes a miro-workflow skill that describes how to assemble a few common board layouts from the atomic create-item tools.
Five layouts are covered:
Workflow
Trigger phrase examples
Sprint Board
"set up a sprint board for sprint 42"
Retrospective
"create a retro board for our team"
Brainstorm
"brainstorm board for [topic]"
Story Map
"map the user journey for [product]"
Kanban
"kanban for our backlog"
Each one specifies frame sizes, gap math, sticky placement, color conventions (the Miro API takes hex for frame fills and named values for stickies; two different vocabularies), and the tool-call order to follow. For single-tool calls ("add one sticky") or read-only inspection, the atomic tools are usually enough on their own.
The skill is new and likely needs adjustment as more layouts get exercised against real boards. Files live under skills/miro-workflow/. Claude Code auto-discovers skills from ~/Projects/claude-code-config/skills/; for other hosts, follow your client's skill installation docs.
The badge above is awarded by SkillCheck Pro; this skill currently scores 92/100 with zero Critical issues and zero Warnings. Skills in this repo earn the badge at ≥90/100 with zero Critical and zero Warnings; below that bar, the badge is dropped until the skill passes again.
Companion CLI: miro-cli
For shell-first usage, this server has a sibling: miro-cli. Same author, same Miro API coverage, different runtime. One verb per endpoint, JSON in and out, plus a local SQLite mirror for offline search. Reach for it when you're scripting Miro in CI, bulk-migrating stickies between boards, or querying boards offline without burning API quota.
The MCP server and the CLI overlap deliberately. Use the MCP server when an agent runtime is in the loop (Claude Code, Claude Desktop, Cursor). Use the CLI when bash, Makefiles, or CI are. Use both if you want.
Install:
brew install olgasafonova/tap/miro-cli
# or
go install github.com/olgasafonova/miro-cli/cmd/miro-cli@latest
Companion MCP Apps server: miro-mcp-apps
If you want Miro data to render as interactive UI in the chat (cards, tables, color clusters, SVG graphs) instead of streamed JSON, there's a TypeScript sibling: miro-mcp-apps. Eleven tools built on the MCP Apps extension (SEP-1865), reusing the same MIRO_ACCESS_TOKEN. The two servers run side-by-side: this one for the 110-tool CRUD surface, that one for visual at-a-glance views.
The MCP Apps pattern is TypeScript-only today (Go SDK has no ext-apps helpers), which is why this companion lives in a separate repo instead of bolting onto the Go server.
Token Efficiency
The full tool surface (110 tools) costs roughly 20.8K tokens of preload; the essentials profile trims that to ~2.8K. For sessions where that footprint matters, set MIRO_TOOLS_PROFILE=essentials in your client config; the server then registers a curated 14-tool subset (boards, list/find/search, sticky/text/frame/connector creation, list/get/update/delete items) plus one discovery meta-tool. Agents reach the rest via miro_tool_search on demand.
Profile
Tools
Preload tokens (est.)
full (default)
110
~20,795
essentials
15
~2,848
Savings: ~17,760 tokens (86.2% reduction) when you opt into essentials. Description tokens are exact (JSON-marshaled); schema cost is estimated at 200 bytes per tool. Reproduce locally with go run ./cmd/token-count/.
miro_tool_search(query?, category?, limit?) is registered in both profiles. It searches tool names, titles, descriptions, and categories with weighted keyword scoring (name 3×, title 2×, category 2.5×, description 1×), returns up to 50 matches, and never recommends itself. Use it when you don't know which tool to reach for, or to scope to a category before browsing. Empty query plus a category returns the category's tools alphabetically.
Miro opened a beta of their official MCP server in December 2025, with around 10 tools, and has grown it substantially since. Comparison refreshed 18-09-2026 by enumerating the 45 tools the hosted server actually registers at mcp.miro.com, rather than counting the documented set — the docs lag the deployment in both directions, and the surface moves fast enough that a count is perishable: 65 on 13-08-2026, 63 on 18-08, 61 on 09-09, 45 on 18-09.
The September drop is a consolidation rather than attrition. Every one of the sixteen removed tools is a content read or write path superseded by the canvas_* family — board_list_items, board_show, context_explore, context_get, all three diagram_*, all three doc_*, all four layout_* — and the server's own connection instructions now lead with Canvas Composer SVG as "Miro's unified interaction/representation protocol for the canvas". The 09-09 pass saw it coming: layout_* were still registered but already self-describing as deprecated in favour of canvas. Earlier churn: board_trash and space_delete, present on 17-08-2026, were gone the next day, leaving the surface with no hard-destructive verbs; all five code_widget_* tools present on 18-08 were gone by 09-09, while canvas_search, canvas_load_format_skill and content_item_list_roles arrived. Recent additions come from their changelogs on new MCP tools and code widget endpoints.
Feature
This Server
Official Miro MCP
Tools
110 (or 15 in essentials profile)
45
Transport
stdio + HTTP
HTTPS only (hosted at mcp.miro.com)
Self-hosting
Yes
No
Offline mode
Yes
No
Auth
Token + OAuth2
OAuth 2.1, dynamic client registration (Enterprise orgs need admin enablement)
Board list metadata
id, owner, team, created/modified timestamps
name and URL
Diagram generation
Mermaid, parsed locally; native diagram items readable via miro_list_diagrams/miro_get_diagram
None since 18-09-2026 (diagram_create_mermaid, diagram_update_mermaid and diagram_get_mermaid_instructions removed; diagrams now go through the canvas SVG tools)
AI context
No
None since 18-09-2026 (context_explore and context_get removed)
None since 18-09-2026 — all four layout_* tools removed, having self-described as deprecated in favour of the canvas tools on 09-09
Canvas as SVG
Read + create + update (data-miro-id diff; local geometry transform; spatial approximation). Reads scope to a single frame via frame_id, and read output is directly re-submittable to the update tool
Yes (read, create, update from SVG). Reads scope to a rectangle or a widget_ids list and refuse areas above 500 widgets; canvas_search (pattern and regex over the SVG) narrows the scope first. Read output is re-escaped, so it cannot be fed back to the update tool verbatim
Spaces & sections
No
Yes (10 tools) — workspace-level grouping of boards, not board content
Where the differences come from. Spaces, sections, canvas-SVG and prototypes have no endpoints in the public REST API v2 — checked 13-08-2026 against the 114-path OpenAPI spec tracked in api-tracking/, and re-checked by probing the live API, since the spec omits endpoints Miro does ship. Those run on Miro-internal APIs. The reverse also holds: mindmaps, tags, groups, connectors and export are public REST surfaces the official server has not wrapped.
Board context used to be a design difference rather than a gap, and as of 18-09-2026 it is neither: context_get returned a server-side AI summary of a board and has been removed along with context_explore. miro_get_board_content returns the structured content itself — items by type, frame hierarchy, connectors and tags — and lets the model already in your session do the summarizing, with no second inference hop.
When to use the official server: You want zero-setup via plugin marketplace, OAuth 2.1 enterprise security, spaces and sections, comments, or SVG/prototype workflows. Board context extraction is no longer among them — those tools were removed on 18-09-2026.
When to use this server: You need full REST coverage (110 tools, or a tunable 15-tool essentials mode), offline/self-hosted operation, richer board listing metadata, bulk ops, mindmaps, tags, connectors, or export.
Both can coexist — use different MCP server names in your config.
Performance
Caching: 2-minute TTL reduces API calls
Rate limiting: Adapts to Miro's rate limit headers
Circuit breaker: Isolates failing endpoints
Parallel bulk ops: Creates items concurrently
Token validation: Fails fast on startup with clear error
Transient error retry: Auto-retries 502/503/504 with backoff
# Build
make build
# Run tests
make test
# Run with coverage
make test-cover
# Lint
make lint
# Build for all platforms
make build-all
# See all targets
make help
Docker Deployment
Quick start:
# Using Docker Compose
export MIRO_ACCESS_TOKEN=your-token
docker compose up -d
# Or build and run manually
docker build -t miro-mcp-server .
docker run -e MIRO_ACCESS_TOKEN=xxx -p 8080:8080 miro-mcp-server
docker-compose.yml features:
Health checks configured
Resource limits template
Environment variable passthrough
Debugging with MCP Inspector
Test and debug the server interactively with MCP Inspector:
# Install and run (no setup required)
npx @modelcontextprotocol/inspector miro-mcp-server
# With environment variables
MIRO_ACCESS_TOKEN=your-token npx @modelcontextprotocol/inspector miro-mcp-server
Four tools need an Enterprise plan and fail with a hint saying so: the three PDF/SVG export tools (miro_create_export_job, miro_get_export_job_status, miro_get_export_job_results) and miro_get_org_audit_logs, which additionally needs the auditlogs:read scope. miro_get_board_picture works on every plan. Everything else works on any plan.
The 6 code widget tools and 5 comment tools use Miro's v2-experimental API; availability may vary by account or plan, and the endpoints may change before GA.
Like This Project?
If this server saved you time, consider giving it a ⭐ on GitHub. It helps others discover the project.