cnvs-whiteboard-skills

Agent Skills + MCP server discovery for cnvs.app — the free, no-signup real-time collaborative whiteboard. Any AI agent can read, write, draw, diagram, and subscribe to live human edits on a shared canvas, either through the hosted MCP server or the REST fallback.

Works with Claude Code, Claude Desktop, Cursor, OpenCode / OpenAI Codex, Aider (with MCP plugin), any other MCP-speaking client, and Hermes-family agents.

MCP Server

Endpoint: https://cnvs.app/mcp (Streamable HTTP, protocol 2025-06-18, no auth — the board ID is the access key).

Registered in the official MCP Registry as app.cnvs/whiteboard.

Capabilities

• Tools (10): open_board, get_board, get_preview, add_text, add_link, add_image, draw_stroke, move, erase, wait_for_update. All ten have 1:1 REST mirrors at https://cnvs.app/api/boards/<id>/… for runtimes that can't speak MCP.
• Resources (2): cnvs://board/{id}/state.json (full snapshot, subscribable) and cnvs://board/{id}/preview.svg (visual render, subscribable).
• Subscriptions: resources/subscribe supported with notifications/resources/updated pushed over SSE, debounced ~3 s after activity settles.
• Live machine-readable manifests: /quotas.json, /openapi.json, /llms.txt, /.well-known/mcp.json, /.well-known/mcp/server.json.

Installation

Claude Desktop / Cursor / any MCP client

One-line config — add this to your client's mcpServers object:

{
  "mcpServers": {
    "cnvs": {
      "type": "http",
      "url": "https://cnvs.app/mcp"
    }
  }
}

Claude Code CLI

claude mcp add --transport http cnvs https://cnvs.app/mcp

REST-only (no MCP client)

# Create a board
curl -X POST https://cnvs.app/api/boards

# Add text
curl -X POST https://cnvs.app/api/boards/<id>/texts \
  -H 'Content-Type: application/json' \
  -d '{"x":100,"y":200,"content":"# Hello","author":"ai:myagent"}'

# Long-poll for live changes
curl "https://cnvs.app/api/boards/<id>/wait?timeout_ms=25000"

Full REST reference: /llms.txt, /openapi.json.

Skills included

Two related Agent Skills live in this repo, published under the agentskills.io open standard so they work in any compatible runtime. Install them into ~/.claude/skills/ to teach the agent how to use the MCP server well (preview-before-JSON, REST-over-MCP for writes, author-tag conventions, subscription-then-react loop).

cnvs-whiteboard/ — PRIMARY

Teaches an AI agent how to collaborate on a cnvs.app board in real time:

• Read the shared canvas state
• Draw / diagram / flowchart / annotate — add, update, move, and delete text, links, sticky notes, strokes, images, and Mermaid diagrams (flowcharts, sequence / class / ER diagrams, mind maps, gantt charts, concept maps)
• Render the SVG preview when the board contains drawings
• Subscribe to live human edits via MCP, react via REST

Activates on any cnvs.app board reference (URL https://cnvs.app/#<id>, cnvs://board/<id>/..., or a bare board ID) or phrases like "collaborate on / draw / diagram / annotate / watch a shared whiteboard or canvas."

mcp-listen/ — HELPER (v0.3.0)

Push-to-model pump for Streamable-HTTP MCP servers with capabilities.resources.subscribe: true. Opens a session, subscribes to the resource URIs the caller hands over, and emits one JSON line per notifications/resources/updated event on stdout — designed to be wrapped by Claude Code's Monitor tool so every server push becomes an in-chat notification (no polling, no log tailing).

Helper-only activation since v0.3. Triggers only when the caller — a user or another skill — has supplied both an MCP server URL and at least one resource URI. Does NOT auto-activate from generic phrasing like "watch this file" or "stay in the loop". cnvs-whiteboard is the canonical caller and delegates its push channel here.

Output split since v0.3. Only the actionable resource_updated event reaches stdout by default. Connection / subscription / heartbeat / transient-error events go to stderr so a wrapping Monitor doesn't fire on the listener's own bookkeeping. Pass --verbose (or -v) to merge everything onto stdout for debugging or single-channel logging (legacy v0.2 behavior).

Registry status. Intentionally not listed in centralised skill registries yet. Current registry pipelines (e.g. majiayu000/claude-skill-registry-core as of 2026-05-03) archive only SKILL.md + generated metadata; they do not mirror bundled scripts/** or package.json, so an entry for mcp-listen would publish a broken skill. It will be (re-)submitted once core registries support directory-style skill archival with bundled-file mirroring and security scanning — see PR majiayu000/claude-skill-registry-core#40 for the directory-archive primitive that just landed and the open question on bundled-script scanning. Until then, install it via curl from cnvs.app (below) or git clone this repo.

Install the skills

# cnvs-whiteboard (SKILL.md only — no deps)
mkdir -p ~/.claude/skills/cnvs-whiteboard && cd ~/.claude/skills/cnvs-whiteboard
curl -O https://cnvs.app/cnvs-whiteboard/SKILL.md

# mcp-listen (SKILL.md + scripts/ + npm install)
mkdir -p ~/.claude/skills/mcp-listen && cd ~/.claude/skills/mcp-listen
curl -O https://cnvs.app/mcp-listen/SKILL.md \
     -O https://cnvs.app/mcp-listen/package.json \
     --create-dirs -o scripts/listen.mjs https://cnvs.app/mcp-listen/scripts/listen.mjs
npm install

Alternatively git clone this repo and cp -r .claude/skills/* ~/.claude/skills/ — the .claude/skills/ directory contains symlinks into cnvs-whiteboard/ and mcp-listen/ for Claude Code's expected layout.

Spec compliance

Both skills follow the agentskills.io spec:

• name matches the containing directory name
• description under 1024 characters, imperative phrasing, explicit trigger keywords
• scripts/ subdirectory for bundled code (mcp-listen)
• license: MIT, compatibility field declared, cross-references in each skill's "Related skills" section to prevent activation overlap

Why two skills

Earlier versions had a single page (/skill-cnvs.md on cnvs.app — now a legacy redirect). Splitting into two coherent units follows the spec's "design coherent units" best practice:

• cnvs-whiteboard owns the cnvs.app-specific workflow (read / write / render / gotchas).
• mcp-listen owns the generic push-to-model pattern (usable against any MCP server).

Cross-references keep them pairing cleanly: cnvs-whiteboard mentions mcp-listen as its delegated push channel; mcp-listen mentions cnvs-whiteboard as the companion for cnvs-specific work. No activation overlap in practice.

Canonical sources

Skills are also served live at:

• https://cnvs.app/cnvs-whiteboard/SKILL.md
• https://cnvs.app/mcp-listen/SKILL.md
• https://cnvs.app/mcp-listen/scripts/listen.mjs
• https://cnvs.app/mcp-listen/package.json

The cnvs.app URLs are the canonical install targets for curl-based installers; this GitHub repo is the reference for aggregators that crawl public sources (skills.sh, etc.) and for users who prefer git clone.

Discovery surfaces

Live

• Official MCP Registry: app.cnvs/whiteboard — browse
• cnvs.app /.well-known/mcp.json — links both skills
• cnvs.app /.well-known/mcp/server.json — MCP Registry entry with publisher-provided _meta.skills[]
• cnvs.app /llms.txt — LLM-friendly full reference
• majiayu000/claude-skill-registry-core — cnvs-whiteboard listed (PR #31 merged 2026-05-04). mcp-listen intentionally held back; see Registry status on the mcp-listen/ section above.
• Lobehub (used by Hermes agent) — both skills imported; MCP server submitted

Planned / under consideration

Notes for future expansion of distribution. Not active yet — kept here so the next maintenance window doesn't have to re-discover the landscape.

• majiayu000/claude-skill-registry-core — re-submit mcp-listen as a separate PR now that PR #40 (directory archives + bundled-script security scanning) is merged. scripts/listen.mjs is even hardcoded into their test assertions, so the pipeline already exercises our exact case.
• daymade/claude-code-skills (vendored model) — fork, copy cnvs-whiteboard/ and mcp-listen/ directly into their tree (their layout already expects scripts//references//assets/), update marketplace.json, single PR for both skills. No bundled-files concern because the files are copied in.
• VoltAgent/awesome-agent-skills (curated awesome-list, 20k ⭐) — markdown link, ≤10-word description. Their CONTRIBUTING explicitly excludes brand-new skills ("give your skill time to mature and gain users before submitting"); revisit once there's documented external usage.
• sickn33/antigravity-awesome-skills (36k ⭐) — manual path: fork, add skills/<name>/SKILL.md, run npm run validate. Their <!-- registry-sync: ... --> header in README suggests automated pull from upstream registries, so a successful entry on majiayu000/claude-skill-registry-core may propagate here without a separate PR — confirm before submitting twice.
• agentskills/agentskills (17.8k ⭐, the spec org behind agentskills.io) — investigate whether they accept reference implementations; mcp-listen as a worked example of MCP resources/subscribe consumption could fit there.
• vercel-labs/agent-skills / vercel-labs/skills — almost certainly curated, but read CONTRIBUTING before ruling out.

Skipped intentionally: anthropics/skills (curated, no community PRs); personal collections under ~100 ⭐; language-specific lists unless a relevant audience emerges.

License

MIT. See LICENSE.