Anchored in the Word
Every answer cited in scripture and historic commentary — no hallucinated verses, no denominational slant.
Our why
Anchor is a study tool for people who already read the Bible and want help going deeper.
“We have this hope as an anchor for the soul, firm and secure. It enters the inner place behind the curtain.”
The problem we're solving
Over the last two years, something quiet has shifted. The people in your pews, your small group, and your own family have started asking ChatGPT, Gemini, Claude, and Meta AI. Those tools were trained to be fluent, not faithful — and on scripture they are getting things wrong in ways most readers cannot detect.
A verse that isn’t in the Bible.
Chatbots make them up — complete with chapter and verse numbers — and deliver them with the same confidence as a real quotation. Most readers can’t catch it before the fake verse ends up in a text, a card, or a sermon.
A quote that blends the cadence of the KJV, the word choice of the ESV, and the phrasing of the NIV into a single sentence nobody ever wrote.
The model does not distinguish between translations. It averages them. The result reads like scripture and isn’t.
“Most Christians believe…”
On questions where Christians have read scripture differently for two thousand years — baptism, communion, eschatology, church governance — general AI picks a winner and states it as fact, with no flag that the matter is contested.
Confident paragraphs about what scripture teaches — with no verses attached, and no way for the reader to check the claim against the text.
When a chatbot doesn’t know, it doesn’t say so. It produces something that sounds authoritative and moves on.
A fourteen-year-old in crisis types a question about faith and gets a confidently wrong answer — no sources, no humility, no off-ramp to a real human.
The tools most likely to be consulted at 2 a.m. were not built with pastoral care in mind. There is no crisis protocol. There is no “talk to someone.”
The people asking those questions don't know the answers are wrong. That is the actual problem. A pastor cannot responsibly recommend any of these tools. Anchor is built to be the one you can.
Meet Anchor
A reader opens anchor.bible and types a question — “What does the Bible say about anxiety?”, “How did the early church read Romans 8?”, “What does the word ‘propitiation’ actually mean?” Anchor searches a library of four complete Bible translations and six classical commentaries, writes an answer in plain English, and shows the reader every verse and commentator it drew from.
If it can't cite a source for a claim, it declines to answer rather than invent one. That single rule — cite or refuse — is the spine of the product.
At a glance
Bible translations indexed
KJV, ASV, WEB, and Young’s Literal — every word ingested from authoritative public-domain source texts, not recalled by a language model.
Classical commentaries & reference works
Matthew Henry, Gill, Spurgeon, Jamieson-Fausset-Brown, Nave’s Topical, and Adam Clarke. Three centuries of expositors, retrievable by passage.
Tradition lenses
Mere Christian, Reformed, Catholic, Orthodox, and Methodist. Each lens changes both which commentators are consulted and how the answer is framed.
Indexed commentary files, book-by-book
Each Gospel, each epistle, each prophet is its own indexed file — so a question on Romans retrieves from Henry on Romans, Gill on Romans, and JFB on Romans.
Eval questions passing before every deploy
An extensive automated adversarial test suite graded against fabrication, refusal, tradition-balance, and crisis-handling. Ships only when every test passes.
Invented verses. Ever. By design.
Every verse reference is pulled from an indexed translation. The model cannot “remember” a verse — only quote one Anchor has already retrieved.
The posture
Anchor's posture is deliberately narrow. It does not pronounce salvation. It does not interpret dreams. It does not diagnose. It does not take sides in contested intra-Christian debates.
Instead, it names the tradition each view belongs to, cites the passages each side relies on, and points the reader back to scripture and their own pastor. It is a tool for going deeper, not a substitute for the shepherd. That line — between what a study tool can do and what a pastor does — is drawn into the system prompt itself, not painted on after.
Anchor is not a replacement for your pastor, your church, or your own study of scripture. It is not a counselor, a prophet, or a spiritual authority. It is a study aid — a quiet companion for people who want to understand what they're reading.
Inside the corpus
The texts Anchor draws on are not summaries, paraphrases, or AI-generated retellings. They are the full source documents, ingested verbatim, indexed for retrieval. No user is ever shown a claim that did not come out of one of the following files.
| Source | Author / Publisher | Scope | Files |
|---|---|---|---|
| Scripture — four complete English translations | |||
| KJV | King James Version | Old + New Testament, complete | 66 books |
| ASV | American Standard Version (1901) | Old + New Testament, complete | 66 books |
| WEB | World English Bible | Old + New Testament, complete | 66 books |
| YLT | Young’s Literal Translation | Old + New Testament, complete | 66 books |
| Reformed & evangelical commentary | |||
| Matthew Henry | Matthew Henry (1706–1714) | Every book of the Bible | 6 volumes |
| John Gill | John Gill (1746–1763) | Every book of the Bible | 66 files |
| Spurgeon — Treasury of David | Charles Spurgeon (1865–1885) | Psalms — all 150 | 3 split files |
| Cross-tradition reference | |||
| Jamieson-Fausset-Brown | Jamieson, Fausset & Brown (1871) | Critical & explanatory, whole Bible | 30 files |
| Nave’s Topical | Orville J. Nave (1896) | ~20,000 topical cross-references | 1 volume |
| Wesleyan / Methodist commentary | |||
| Adam Clarke | Adam Clarke (1810–1826) | Every book of the Bible | 66 files |
Every file is copyright-clean, ingested from authoritative public-domain editions, and stored as structured text — not scraped, not AI-summarised. A citation to Matthew Henry on Romans 8 points at the actual paragraphs Matthew Henry wrote, not a model's recollection of them. See the full Sources page for a complete list.
How an answer is produced
Most AI Bible tools are generation-first — the model writes whatever it thinks is true, and the verses are decorative. Anchor inverts that. Every answer begins with a grounded retrieval against the indexed corpus; the language model is constrained to speak from those retrieved passages.
Default (Mere Christian), Reformed, Catholic, Orthodox, or Methodist / Wesleyan. Each lens selects a different system prompt and a different set of retrieval stores — so a Methodist reader and a Reformed reader querying the same verse are drawing on different commentators, on purpose.
Gemini 2.5 Flash is invoked with the fileSearch tool pointed at up to five File Search stores (Bible translations + the commentary stores assigned to that lens). The model is literally prevented from answering without first retrieving from them.
The system prompt (different for each lens) enforces a non-negotiable rule: every factual claim about scripture or a commentator must be inline-cited in [Book Chapter:Verse] format. If a claim cannot be cited, the instruction is to stop and say so — not to guess.
As Gemini streams the answer back, its groundingMetadata carries the exact chunks it drew from. A small resolver converts internal file names into human-readable citations — “Matthew Henry on Romans”, “Spurgeon, Treasury of David on Psalms”, “Genesis (ASV)”.
The answer streams in token-by-token; the full list of sources appears below it, deduplicated and labeled. Every answer is a claim and a bibliography. Nothing is shown without showing where it came from.
What “cited” actually looks like
This is the shape of a real Anchor response. The citations in brackets render as clickable pills in the product; the sources panel shows the exact commentaries that were retrieved.
Question
“What does the Bible say about anxiety?”
Scripture speaks to anxiety directly and often. Paul writes — “Be careful for nothing; but in every thing by prayer and supplication with thanksgiving let your requests be made known unto God” (KJV) — setting prayer and gratitude as the posture that guards the heart. Jesus, in the Sermon on the Mount , frames anxious thought as a question of trust in the Father's provision. Peter carries the same note: , “casting all your care upon him; for he careth for you.”
Matthew Henry, writing on Philippians 4, notes that Paul's instruction is not to feel no concern, but to convert concern into prayer. “The way to be careful for nothing, is in everything to make our requests known to God.” Matthew Henry on Philippians 4.
The five lenses
Anchor does not flatten denominational difference into a generic “what Christians believe.” A reader picks a lens, and that lens determines both the commentators consulted and the posture of the answer. Each lens is labeled — nothing is hidden behind a default.
Historic creedal Christianity — Apostles’ Creed, Nicene Creed. Presents the range of views on contested questions and refuses to pick a winner.
Draws on
All commentary stores — Matthew Henry, Gill, Spurgeon, JFB, Nave’s, Adam Clarke.
Reformed framing on scripture’s authority and soteriology. Weights the great Reformed expositors.
Draws on
Matthew Henry, John Gill, Spurgeon’s Treasury of David, JFB, Nave’s.
Sacramental and Magisterial posture. Draws explicitly on the Church Fathers — Augustine, Aquinas, Chrysostom — as named in the answer.
Draws on
Jamieson-Fausset-Brown, Nave’s Topical Bible.
Eastern Orthodox reading — Chrysostom, Basil, the Philokalia — with emphasis on theosis and Holy Tradition.
Draws on
Jamieson-Fausset-Brown, Nave’s Topical Bible.
Wesleyan perspective via Adam Clarke’s commentary and Wesley’s sermons. Emphasis on prevenient grace and sanctification.
Draws on
Adam Clarke’s Commentary, JFB, Nave’s.
When a question lands in a genuinely contested area — baptism, communion, predestination, church governance, women in ministry, eschatology — the Mere Christian lens will acknowledge the disagreement, name each tradition, cite the passages each side uses, and decline to pick a winner. It will then point the reader back to scripture and to their own pastor. That is a product rule, not a marketing promise.
The discipline
The hardest part of building a trustworthy Bible AI is not what it says. It's what it refuses to say. Anchor ships with three kinds of refusal built into the system prompt. Every one is intentional.
On these, Anchor will gently but firmly refuse, and explain why.
If a user mentions self-harm, suicide, or any indication of crisis, Anchor immediately shows local crisis resources — Lifeline, Samaritans, 988, local emergency services — before anything else. Scripture is only offered afterward, if at all.
This branch takes precedence over every other instruction.
For pastoral counseling, major life decisions, or doctrinal disputes, Anchor will share the relevant scripture — and then explicitly decline to make the decision or give counsel. The phrasing is consistent: “This is the kind of question a pastor or trusted friend can walk with you through. Here's what scripture says on the topic…”
When a question touches an area where Christians have read scripture differently — baptism mode, communion, eschatology, governance, sign gifts, soteriology — the default lens is required to present the range of major historic views, name each tradition, cite the relevant passages, and refuse to declare a winner.
It is not hedging. It is the product honoring that pastors, not a chatbot, settle contested questions inside a church.
How the promises are kept
“No invented verses” and “every claim cited” are not aspirations — they're testable properties. Anchor ships with an extensive automated evaluation harness that runs the product through adversarial questions, grades each answer against expected behaviours, and currently passes at 100%.
The harness tests the real categories of failure a pastor would worry about:
If a change to the codebase regresses any of those, the eval fails loudly before the new version reaches a reader. The harness is re-run on every material model or prompt change.
Where the care shows
None of the following is marketing copy. Each is a real engineering choice made in the course of building Anchor, and each is the kind of thing a careful reader notices.
Spurgeon’s commentary on the Psalms is 12 MB — large enough to strain Gemini’s re-indexing pipeline. It was split into psalms-001-050, -051-100, and -101-150 for reliability. The citation resolver strips the range suffix before display, so a reader always sees “Spurgeon, Treasury of David on Psalms” — never the internal file name. The mechanics stay invisible.
The free-tier cap is enforced server-side in Postgres using a SHA-256 hash of the client IP combined with a server-side pepper. The raw IP is never stored. If the database gate itself errors, Anchor fails open — a momentary outage never silences a believer’s question.
There is no single “the prompt.” Each tradition lens owns its own file (default.ts, reformed.ts, catholic.ts, orthodox.ts, methodist.ts) with a PROMPT_VERSION constant. Prompts can be changed intentionally, reviewed, and rolled back — they can’t drift silently.
Every text/background pair in the visual brand kit (Navy on Cream, Charcoal on Cream, Teal on Cream, Cream on Navy, Gold on Navy) is validated against WCAG AA. Gold on Cream fails at 2.6:1 — so Gold is never used for body text, only as a decorative accent. A reader with low-vision concerns will not bounce off the page.
Answer reveals fade gently. Citations appear a heartbeat after the answer has settled. For readers with vestibular sensitivity or reduced-motion preferences set in their OS, every animation collapses to an instant opacity change. Scripture should not induce motion sickness.
The brand kit explicitly bans unlock, supercharge, ultimate, secret, revolutionary, and Christianese marketing language (anointed, blessed). Scripture doesn’t need emojis either. The tone is wise mentor, never excited marketer.
Gemini’s re-indexing pipeline has occasional flaky windows. Anchor ships with a retry script that re-runs failed ingestions on a cadence until the store reaches steady state, logging every attempt. A failed upload never quietly turns into a missing citation.
The pastoral commitments
If Anchor cannot find a specific verse or commentator to back a claim, it says so and stops. The model is forbidden from filling the gap.
Every verse reference is pulled from an indexed translation. The generator cannot “remember” a verse — it can only quote from a retrieved one.
On intra-Christian disagreements, Anchor presents traditions by name, cites what each relies on, and sends the reader back to scripture and to their pastor.
Anchor refuses to pronounce on a person’s salvation, interpret their dreams, or deliver “a word from the Lord.” On pastoral matters, it points to pastors.
If a reader indicates self-harm or crisis, resources appear first. Scripture is offered afterward, only if appropriate.
Who made this
Anchor is built and maintained by a small independent team in Aotearoa New Zealand. It is not affiliated with any church, denomination, or seminary.
Anchor exists because the general-purpose chatbots are getting the Bible wrong — inventing verses, flattening theology, ignoring tradition, giving pastoral counsel they have no right to give. A pastor cannot responsibly recommend any of them. Anchor is built to be the one you can.
Ask it anything. Read the answers. Check the citations against your own shelf. If something doesn't hold up, tell us — the discipline of this product is your feedback loop, not our assumption.
Every answer cited. Every source named.