AI connector
MCP Integrations
Holy Grail exposes one production remote MCP server for AI clients. The connector is public by default, uses OAuth for account-scoped access, and separates read tools, write tools, publication tools, and destructive account controls through explicit scopes and confirmations.
https://holygrailjournal.com/mcp. Support: support@holygrailjournal.com.
Canonical endpoint
Use the apex connector URL only. Do not configure AI clients with the www host.
The production MCP endpoint is `https://holygrailjournal.com/mcp`. OAuth and protected-resource metadata are served from the same canonical origin.
- MCP endpoint: https://holygrailjournal.com/mcp
- Protected-resource metadata: /.well-known/oauth-protected-resource/mcp
- Authorization-server metadata: /.well-known/oauth-authorization-server
- OpenAI connector listing: /.well-known/openai/connector-listing.json
ChatGPT and OpenAI
Use ChatGPT Developer Mode to create and test the connector, then submit through the OpenAI Apps SDK review flow.
The repo-tracked submission bundle lives under `docs/openai/`, and the live service serves the same listing and registry through `/.well-known/openai/`.
- Connector URL: https://holygrailjournal.com/mcp
- Primary logo: /mcp.png
- Review prompts: docs/integrations/test-prompts.md
- Provider readiness audit: docs/integrations/provider-readiness-audit.md
- External smoke runbook: docs/openai/external-install-smoke-test.md
Claude
Claude custom connectors should use the remote MCP URL directly. Claude Code can install the same endpoint with its HTTP transport, and Anthropic API calls can pass the same URL in `mcp_servers`.
Use a disposable smoke account for write tests and avoid destructive tools except in the dedicated data-control runbook.
- Claude custom connector URL: https://holygrailjournal.com/mcp
- Claude Code: claude mcp add --transport http holygrail https://holygrailjournal.com/mcp
- Anthropic API: mcp_servers[].url=https://holygrailjournal.com/mcp, authorization_token=<HOLYGRAIL_MCP_ACCESS_TOKEN>, tools[].type=mcp_toolset
- First tools to test: connector_self_test, diagnostics_get_connection_status, get_contribution_guide, list_submissions
Gemini CLI
Gemini CLI should be configured with an `httpUrl` MCP server entry. Start with a read-only includeTools profile, then add write tools after OAuth, scope filtering, and replay behavior are verified.
- Server name: holygrail
- httpUrl: https://holygrailjournal.com/mcp
- Recommended default: trust=false
- Use includeTools for read-only and smoke-test profiles
xAI and Grok
xAI Remote MCP Tools use API-level tool configuration. Provide `server_url`, `server_label`, an OAuth bearer value in `authorization`, and a narrow `allowed_tools` list instead of injecting the entire tool surface by default.
- server_url: https://holygrailjournal.com/mcp
- server_label: holygrail
- authorization: Bearer <HOLYGRAIL_MCP_ACCESS_TOKEN>
- Start with diagnostics and read tools
- Do not include destructive tools in default examples
Public-by-default warning
Holy Grail is not confidential storage. Finalized submissions may become public graph content, and even drafts should be treated as unsuitable for secrets or regulated/private data.
Use exact confirmation text and stable idempotency keys for export, deletion, revocation, account closure, and publication workflows.