Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deaf Data Lab

Explore aggregate statistics, understand definitions, compare supported populations and survey periods, and share findings with uncertainty and reproducible context. No Bun, Python, local Census data or access to the application repository is needed.

Version: 0.2.1. Published by William Albright under MIT. Repository: https://github.com/ewilliam/deaf-data-plugin. Marketplace: deaf-data-lab. MCP endpoint: https://deaf-data.vercel.app/mcp. Access: anonymous, approved public aggregates only. Support: https://github.com/ewilliam/deaf-data-plugin/issues. Tested client versions and file delivery: Codex CLI 0.158.0 manifest/schema validation and Claude Code 2.1.274 isolated install/uninstall checks passed. Hosted MCP tools, resources and CSV/JSON export bytes were verified with the official SDK. Full client conversation, upgrade, attachment-delivery and Deaf-user comprehension evaluations remain pending; no claim of completed acceptance is made. Until your client delivers an actual attachment, use the documented usable-text and canonical-query fallback.

Choose your client

All packages include the same three skills and shared interpretation reference. Packaging support is distinct from completed live client acceptance; see the validation status above.

Client Installation Data connection
Codex GitHub marketplace below Bundled remote MCP configuration
Claude Code GitHub marketplace below Bundled remote MCP configuration
Claude web/Desktop/Cowork Upload the claude.zip release asset Review/connect the bundled remote connector; if needed add the endpoint in Connectors
ChatGPT Upload the chatgpt-skills.zip release asset Register the MCP app separately and select it with the skills; see connected-package instructions below

Download the versioned ZIPs from Releases. Names include the version, for example deaf-data-0.2.1-claude.zip. Do not upload GitHub's Source code ZIP or Code → Download ZIP as a plugin: those contain a marketplace repository, with the actual plugin nested inside it.

Codex and Claude Code

In Codex, add this catalog:

codex plugin marketplace add ewilliam/deaf-data-plugin

Open the Codex plugin directory, select the deaf-data-lab marketplace and install deaf-data. Start a new chat after installation. In Claude Code:

claude plugin marketplace add ewilliam/deaf-data-plugin
claude plugin install deaf-data@deaf-data-lab

End installation with your first question:

What can I learn about deaf employment in my state?

The assistant checks available releases and measures, briefly explains coverage, and asks for your state if it is not already known. If discovery is unavailable, a successful connection does not mean there is data to query. Follow the recovery message or contact support with the request ID.

Try “What does the denominator mean?”, “What about Texas?”, “Compare that with hearing adults,” “Use the previous period,” or “Give me a copy-ready finding and a complete CSV.” The assistant carries forward established selections and asks before changing the analysis to obtain coverage. “Latest” means the newest approved, queryable survey period supporting your analysis, not a recent website deployment. The answer names the actual survey window.

Claude web, Desktop and Cowork

Open Customize → Browse plugins (or the Plugins directory in your client), choose the option to upload a custom plugin, and select deaf-data-0.2.1-claude.zip. Install/enable it and review its connector setup. The server is https://deaf-data.vercel.app/mcp and needs no authentication. If the connector is not imported automatically, add that URL as a custom remote connector and enable it alongside the plugin. Start a new conversation and check that all three skills and the data connection are available. Account/workspace policies may restrict custom uploads or connectors.

ChatGPT: skills and data together

Upload deaf-data-0.2.1-chatgpt-skills.zip using Plugins → Add → Upload. This is a standalone skill package with the manifest and skills/ at its root. It intentionally contains no raw MCP configuration; a skills upload does not register the service.

In Create MCP App, register https://deaf-data.vercel.app/mcp as Deaf Data Lab, with no authentication. Install that connection and the skill package and select both in a new Work chat. If developer mode is required, enable it in Settings → Security and login. UI labels and access vary by account. Confirm that the tools appear before asking for statistics.

For a single package that declares your registered connection, copy its technical ID from the connection's browser URL (starts with plugin_asdk_app). A maintainer can then run:

python3 build-archives.py --output connected-release --chatgpt-app-id YOUR_REGISTERED_APP_ID

Upload the resulting deaf-data-0.2.1-chatgpt-connected.zip. The .app.json mapping binds the three skills to that registered connection. The ID must belong to a connection available to your account/workspace; it does not grant access. This account-specific archive is not included in the generic public release. Do not substitute a fabricated ID. Upload acceptance and tool availability still need verification in your ChatGPT account.

For public ChatGPT directory distribution, use the OpenAI submission portal's With MCP path, submit https://deaf-data.vercel.app/mcp, and include the skills from the skills archive in the same draft. Complete domain verification, listing and review requirements. A GitHub release is not directory approval, and a registered personal app ID is not a substitute for submitting the endpoint.

Reproduce upload packages

Maintainers can run python3 build-archives.py --output release-assets from this repository root. Python's standard library creates deterministic ZIPs and SHA256SUMS; end users only download and upload the assets. portable.zip additionally provides Agent Plugins 1.0 root manifests for hosts supporting that format. Claude's legacy HTTP configuration is retained separately.

Packaging references: OpenAI, OpenAI MCP and skills submission, Claude plugins.

Read and share results

Estimates describe the service's defined survey populations and universes. They are not individual records. Answers retain uncertainty and limitations; suppression means withheld, not zero, and missing coverage is different from suppression. Overlapping survey windows and changed geography boundaries can prevent comparison. The service's returned reasons govern interpretation.

Routine answers use plain Markdown with the finding, uncertainty, essential caveats and a source link. They do not rely on HTML disclosure panels or append technical query dumps. Ask “Show the methods and exact query” for sample details, versioned provenance and canonical reproduction. Follow-up answers carry forward your selections without repeating unchanged technical details.

Exports are complete bounded CSV/JSON resources. Where your client supports attachments, the assistant checks completeness, byte count and SHA-256 and retains query/provenance details beside the file. Otherwise it provides usable text and canonical reproduction and explains the file limitation. An embedded resource URI or base64 text is not a download. Exact application deep links are not supported; use the canonical query or export to reproduce an analysis.

Troubleshooting and versions

  • No matching coverage: inspect available shapes; choose an alternative only after understanding its changed population, period, geography or denominator.
  • Suppressed or incompatible: retain the returned reason. Repeating requests will not reveal a value.
  • 403: check that the installed endpoint matches the canonical host; contact support.
  • 429 or busy: respect Retry-After and retry the same request later.
  • Unavailable data: retry the same selection; send support the request ID. Do not replace it with zero.
  • Oversized export: choose a smaller scope explicitly. Never treat a preview as a complete file.

Refresh the marketplace using your client's plugin manager, upgrade deaf-data, then start a new chat. Uninstall it through the same manager (Claude: claude plugin uninstall deaf-data@deaf-data-lab). For rollback, obtain an earlier reviewed vVERSION tag from the public repository, register that local checkout as a marketplace, and reinstall its version. The service preserves tool contracts; breaking changes require a documented matching bundle version. Contact support for supported versions.

About

MIT-licensed Deaf Data Lab plugin for Codex and Claude Code: discover, interpret, compare, and export aggregate statistics.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages