Bible study and sermon preparation grounded in primary sources — original languages, commentary, historical background, manuscripts, and denominational doctrine.
This repo contains two things that work together:
- 11 skills that teach an AI how to do biblical research well — which tool answers which question, how to cite, and when to admit the sources came up empty.
- MCP configuration for the OpenLumin server, which provides the 30 research tools the skills use.
You need both. The MCP server alone gives an AI thirty tools and no judgement about how to sequence them; the skills alone have nothing to call.
- An OpenLumin account in an organisation on the Church plan — openlumin.com
- Claude (claude.ai, desktop, or mobile), Claude Code, and/or Codex CLI
That's it. There are no API tokens to create or paste — connecting is a sign-in. Every connection you approve shows up in OpenLumin → Settings → API tokens, and revoking it there disconnects that client.
The server URL is the same everywhere:
https://api.openlumin.com/mcp/openlumin/mcp
Settings → Connectors → Add custom connector → paste the URL above → Connect. Your browser opens OpenLumin — sign in, click Approve, done. The connection appears in Settings → API tokens on OpenLumin, and revoking it there disconnects the client.
claude mcp add --transport http openlumin https://api.openlumin.com/mcp/openlumin/mcpAdd -s user to make it available in every project rather than only the current one. Then run /mcp inside Claude Code and choose Authenticate — the browser opens the same sign-in + Approve page.
codex mcp add openlumin --url https://api.openlumin.com/mcp/openlumin/mcp
codex mcp login openlumincodex mcp login opens the browser for the same sign-in + Approve flow, then stores the credential locally. Confirm with codex mcp list.
Verification is the client's own check: the connector shows Connected on claude.ai, /mcp in Claude Code, codex mcp list in Codex. A working setup lists 30 tools.
A skill is a directory containing a SKILL.md. Installing one means copying it where your client looks. There is no build step and no registry.
git clone https://github.com/kalib-code/openlumin-skills.git
cd openlumin-skillsAvailable in every project:
cp -r skills/* ~/.claude/skills/One project only:
mkdir -p .claude/skills
cp -r /path/to/openlumin-skills/skills/* .claude/skills/cp -r skills/* ~/.codex/skills/Restart your client, or start a new session. Skills load automatically when a request matches one.
This repo is a Claude Code plugin marketplace, so two commands bring the skills and the MCP configuration together — skip Part 1 and the copying above entirely:
/plugin marketplace add kalib-code/openlumin-skills
/plugin install openlumin@openlumin
On first use, authenticate via /mcp (browser sign-in + Approve). Nothing else to configure.
Ask something that requires evidence:
What does the Greek word behind "propitiation" in Romans 3:25 mean?
You should see the model select word-study, call OpenLumin tools, and answer with citations. If it answers fluently without calling a tool, the skills are not loaded — check the install path and restart.
Each is selected by the kind of input you give it, not by subject. That is what keeps eleven skills from competing for the same request.
| Skill | Give it | Answers |
|---|---|---|
passage-exegesis |
a passage reference | What does this text mean? |
word-study |
a Greek or Hebrew word | What does this word mean, and how is it used? |
topical-study |
a theme | What does the Bible say about this? |
biblical-backgrounds |
a worldview question | What would the original audience have assumed? |
character-study |
a person's name | Who was this, and where do they appear? |
textual-criticism |
a disputed verse | Why do translations differ here? |
sermon-preparation |
a passage to preach | Research toward Sunday |
small-group-guide |
a passage or theme | Discussion questions and leader notes |
theological-review |
an existing draft | Is this doctrinally sound? |
sacred-test |
a proposed use of AI | Should we use AI for this at all? |
bible-evidence-grounding |
— | Shared discipline the other ten rely on |
Two behaviours matter more than any feature.
sermon-preparation asks whether you have read the passage and prayed before it offers research. The first encounter with a text belongs to the pastor, the Word, and the Spirit — not to a research tool. AI is extraordinary at research and incapable of revelation, and the skills hold that line rather than blurring it.
theological-review stops rather than warns. When a draft contradicts your denomination's stated doctrine, it cites the specific statement and declines to continue until you address it. A caution you can scroll past is not a guardrail. If you genuinely dissent on a secondary matter you can override — deliberately, which is the point.
The corpus is uneven, and some gaps are invisible in the output — a query can return a confident-looking answer built on a single row. The skills are written to disclose these rather than paper over them.
- The Hebrew lexicon has holes. hesed (H2617) is absent entirely, as is H2616. Worse, H2618 is a decoy — same spelling, same transliteration, but it is Hesed, a man, a commissary officer under Solomon (1 Kgs 4:10). H6960 ("wait on the Lord") is likewise missing while its neighbours resolve. The skills are instructed to report an empty result rather than reach for an adjacent number.
- Strong's numbers must be zero-padded —
H0026, neverH26. - Place data is thin. 118 places, with verse links covering only Genesis and Exodus — so a query for Jerusalem returns one verse and looks like a real answer.
- Name indexing is inconsistent. Paul is filed as "Saul", Peter as "Simon", Timothy as "Timotheus". Searches match name or description and are unordered, so a low result limit can hide the person you asked about.
- Denominational coverage is eight bodies: PCA, Assemblies of God, Foursquare, Church of God, Non-denominational Evangelical, Southern Baptist Convention, Vineyard, and American Baptist Churches. Lutheran, Methodist, Roman Catholic, Anglican, Nazarene, Wesleyan, Christian Reformed and Reformed Baptist return nothing.
- Manuscript data names editions, not witnesses. The textual-evidence tool returns versional corpora (Leningrad, Peshitta, Clementine Vulgate, SBLGNT) and refuses cross-witness alignment. It cannot tell you which manuscripts omit a verse or how old they are, so the skill routes that question through translation comparison and study notes instead of inventing manuscript sigla.
Full detail, with the queries that expose each gap, is in skills/bible-evidence-grounding/references/data-gaps.md.
| Symptom | Cause | Fix |
|---|---|---|
| Sign-in page appears but shows "available on the Church plan" | You are signed in, but your organisation is not on the Church plan — the consent page gates before connecting so the client never half-connects | Upgrade at openlumin.com/settings/billing, then Connect again |
| Browser opened but the client still says unauthenticated | The approval page expired (requests live 10 minutes) or the tab was closed before Approve | Start the connection again from the client |
| Disconnected a client, want it gone everywhere | Each OAuth connection is a token in OpenLumin → Settings → API tokens | Revoke it there — the client loses access on its next request |
Missing or invalid OpenLumin token (401) |
The stored connection was revoked in Settings | Reconnect (sign in + Approve again) |
available on the Church plan (403) |
Auth succeeded — the organisation is not on the Church plan | Upgrade at openlumin.com/settings/billing |
| Tool list empty, or server won't connect | Config not loaded | Restart the client; claude mcp list / codex mcp list |
| Answers arrive with no citations | Skills not loaded | Confirm the SKILL.md files are in the skills directory, then restart |
| 403 persists right after upgrading | Entitlement is cached for five minutes | Wait five minutes and retry |
The 401-versus-403 distinction is the one worth remembering: a 403 means your token worked. Only the plan is missing.
MIT