display.dev skill

Public agent skill that teaches your AI assistant how to publish small inline values or large generated files, copy, browse, search, read, edit, share, create an account, sign in, claim an anonymous publish, and iterate from comments on display.dev.

Install

Any agent – skill only

npx skills add display-dev/skill --skill display-dev

Claude Code – plugin (skill + Visualize + bundled MCP server)

/plugin marketplace add display-dev/skill
/plugin install display-dev@display-dev

Then run /mcp and complete OAuth for plugin:display-dev:display-dev. The Claude Code plugin contains:

• the display.dev skill;
• the Visualize skill for on-brand HTML reports, plans, and decks (a pinned release, see claude/visualize.lock);
• the remote MCP server (https://api.display.dev/v1/mcp);
• /display-dev:feedback: lists your artifacts with open comments, shows a plan for each thread, and after you approve it, edits the artifact, replies, and resolves the threads;
• a hook that scans a publish for secrets (API keys, tokens, private keys). It blocks a public publish and asks before any other publish;
• a hook that tells you at session start which of your artifacts have new comments. It reads only the open-thread counts, not the comment text.

The hooks need the dsp CLI for the comment note and run in Claude Code and in Cowork on your own computer. Claude.ai loads the skills and the command but not the hooks. There, connect the display.dev connector for the tools. If you also turned on the display.dev connector in Claude Code, the agent sees two sets of display.dev tools and uses one of them.

Codex – plugin (skill + bundled MCP server)

codex plugin marketplace add display-dev/skill

Then open Codex /plugins, install display.dev, and complete MCP OAuth when prompted. The Codex plugin bundles the skill and the remote MCP server (https://api.display.dev/v1/mcp), so publish, create_upload, make_copy, list, search, read, edit, get_metadata, get_referral_overview, sharing, and comment tools are available after sign-in – no separate MCP setup.

Works across Claude Code, Cursor, Codex, OpenCode, Hermes, and Pi.

What you get

Once installed, your assistant picks up the skill on phrasings like:

• "publish this"
• "share this with the org"
• "post this online"
• "make a private link"
• "share with [email]"
• "publish a report" / "share a dashboard" / "publish Markdown"
• "publish the large HTML file generated in code execution"
• "list my artifacts" / "find the quarterly report"
• "read this artifact" / "search inside this artifact"
• "change this passage in the published report"
• "make a copy of this artifact" / "duplicate version 3"
• "add a company email domain"
• "transfer an email domain"
• "change the email domains add-on"
• "show my referral link" / "show referral rewards"

You can publish without a display.dev account or any setup – you get a 30-day preview URL and a browser claim URL. To create an account or sign in, connect the bundled remote MCP server through OAuth where the host supports it, or run dsp login --email you@example.com in your own terminal and enter the six-digit code there. The installed dsp CLI stores the resulting session. The skill never asks for, reads, or relays the code, and never instructs an agent to inspect your inbox.

Authenticated remote MCP uses inline publish(content=...) for small generated values. For an existing or large HTML/Markdown file, the skill can call create_upload, transfer the raw bytes through api.display.dev, and finalize through the same publish action. The temporary credential may appear in the initiating execution trace but is omitted from final/shared output. The capability lasts 15 minutes and the ordinary plan, visibility, sharing, and version checks still apply. Local stdio keeps its file_path workflow.

Authenticated publishes are private by default for user-scoped credentials and company-visible by default for organization-scoped service keys. The skill passes Company explicitly when the user asks to share with their organization, and it uses Make a copy as the recovery when a service-created artifact cannot become Private.

Claude Cowork Team and Enterprise administrators must allow api.display.dev for code execution and start a new task. Personal Claude Pro and Max cannot currently add that domain, so large Cowork files use the local CLI or dashboard instead; small inline publishing remains available.

Signing in authenticates the CLI; it does not silently transfer an earlier anonymous publish. Open that publish's retained browser claim URL to preserve its existing artifact URL and choose or create the destination organization.

For existing artifacts, the skill uses get_metadata, scoped search, and bounded read to establish the current source context. One exact replacement uses edit, which publishes a new version at the same URL with optimistic concurrency. Complete-source export remains available only through the local CLI for workflows that genuinely need the whole file.

The skill also covers email-domain list, add, DNS verification, transfer status, and cancellation through MCP or dsp email-domains. A current target Owner completes an eligible transfer in Settings → Email Domains; agents and the CLI never perform the irreversible confirmation.

The paid email domains add-on is also managed by a current Owner in Settings → Email Domains. MCP and the direct CLI do not enable or disable it.

For the referral program, the skill uses the read-only get_referral_overview MCP tool to return the link available to the current member or organization connection and to explain earned, available, and used organization rewards. It reports missing links without guessing why they are unavailable.

For an existing artifact, the skill prefers MCP make_copy or dsp make-copy to create a new artifact from a retained version. Discussions and people invited to the source are not carried over. The user can choose recipients for the new artifact.

If an organization blocks public artifacts or outside sharing, the skill explains the returned policy denial and asks before changing the requested audience or recipients. Only an Owner can change either control in the dashboard; the CLI and MCP tools do not provide a policy override.

Codex local development (maintainers)

The Codex plugin bundle lives under codex/display-dev/ and is built from the canonical skill. After editing display-dev/, regenerate the mounts and install from the repo-local marketplace (.agents/plugins/marketplace.json):

bin/sync-mounts.sh
codex plugin marketplace add ./path/to/display-dev-skill

Restart Codex, then install display.dev from the local marketplace via /plugins.

MCP transport – installed stdio fallback

The Codex plugin bundles the remote MCP server by default. For CI, local files, or power-user setups you can run the stdio MCP server instead, in your Codex config.toml:

[mcp_servers.display-dev]
command = "dsp"
args = ["mcp"]

Install the official CLI only with user approval. Skill helpers never download a runtime CLI automatically.

Documentation

Full docs at display.dev/docs/skill.

License

MIT – see LICENSE. Bundles jq 1.7.1 (MIT, © 2012 Stephen Dolan); full text in display-dev/bin/jq.LICENSE.