How to use FaithGPT inside Claude Code, Cowork, and Claude.ai
TL;DR: Add https://api.faithgpt.io/mcp to Claude Code with a single command and sign in through FaithGPT's OAuth consent screen, and Claude gains five tools for Scripture lookup, topical verse finding, grounded answers, doctrine review, and devotional writing. The same server URL works in Claude.ai, Claude Desktop, and Cowork through their connector settings.
A model that has read the entire Bible will still misquote it.
You are three files into a church website, writing the copy for a giving page, and you ask Claude for Romans 8:28 in the King James. What comes back reads right. "And we know that all things work together for good to them that love God, to them who are the called according to his purpose." Maybe that is the verse. Maybe a word moved. You are not going to catch it, because a near-miss on a familiar verse sounds exactly like the verse, and the person reviewing your pull request is checking the button colors.
This is the failure that makes AI Scripture quoting risky in a way that AI code is not. Broken code fails loudly. A misquoted verse ships, gets printed on a bulletin, and sits there.
FaithGPT runs a remote MCP server that closes that gap. Once it is connected, Claude stops answering Bible questions from memory and starts calling a tool that reads FaithGPT's own Bible database. The verse text comes back with its reference attached. You can wire it into Claude Code in one command.
What the FaithGPT MCP server actually is
MCP is the Model Context Protocol, the standard Anthropic uses to let Claude talk to outside services. A remote MCP server is a URL that Claude can call, and FaithGPT's lives at:
It speaks stateless Streamable HTTP over JSON-RPC 2.0. Authorization is OAuth 2.0 authorization code with PKCE, with dynamic client registration and RFC 9728 protected-resource discovery, which is the boring way of saying you never handle a key. Your client registers itself, your browser opens FaithGPT's own sign-in and consent screen, you approve openid profile email faithgpt:mcp, and the token belongs to your account.
Five tools come across:
lookup_scripture takes a reference like John 3:16, Psalm 23, or Romans 8:28-39 and returns the passage text in KJV, ASV, or YLT. KJV is the default. This is the retrieval tool, and it is the one that fixes the misquote problem.
find_scripture_by_topic goes the other direction. Give it a theme, a situation, or a keyword and it returns relevant passages with a short explanation. You can scope it to the Old Testament, the New, or both, and cap it between one and ten results.
ask_faithgpt sends a full question into FaithGPT and returns a Scripture-rooted answer with citations. Use it when the question needs an answer rather than a verse.
review_doctrine takes a block of text, up to 20,000 characters, and reviews it for theological soundness and responsible use of Scripture. FaithGPT is explicit that this is guidance and not qualified pastoral counsel, which is the right line for a tool to hold.
create_devotional builds a devotional around a theme, optionally anchored to a verse, with a length and an audience you pick. It returns a title, a passage, a reflection, and a prayer. It does not publish or save anything.
The part worth reading twice: this connector cannot read or change your private FaithGPT data. Journals, prayers, notes, memories, conversations, media, your subscription, your settings, all off limits. It reads Scripture and generates content. Access is revocable whenever you want it gone.
Claude Code
One command:
Then confirm it landed:
The first time Claude actually calls one of the tools, the FaithGPT consent flow opens in your browser. You sign in and approve the scopes, and that is the whole setup. Nothing gets pasted into a config file, and no API key exists to leak.
To check the connection on its own:
You are looking for a status of Connected. Inside a session, /mcp shows you the servers Claude Code has and lets you handle authentication if a server is sitting there unauthorized.
After that, you ask in English. "Look up Hebrews 12:1-2 in KJV and put it in the hero section." Claude picks lookup_scripture on its own, because the tool descriptions tell it when each one applies. You do not name tools.
Picking a scope
The --scope flag decides who gets the server.
--scope user is what the command above uses, and it is the right default for a solo developer. The server follows you into every project on that machine, so you set it up once.
--scope local keeps it to the current project and to you.
--scope project writes the server into a .mcp.json file at the root of the repository:
Commit that file and everyone who clones the repo gets the server offered to them. Their Claude Code will ask them to approve it before it connects, and each person still signs into their own FaithGPT account, so nobody inherits your credits or your identity. For a team building a church product, this is how you make verified Scripture the default instead of a thing one developer remembers to do.
Claude.ai and Claude Desktop
Custom connectors on claude.ai use the same MCP server URL. In your settings there is a connectors section with an option to add a custom connector, where you paste:
I am hedging the exact menu wording on purpose. Anthropic moves these labels between releases, and the honest instruction is to open your settings and look for connectors rather than to trust a screenshot from whenever this was written. What does not move is the URL and what happens after you submit it: FaithGPT's sign-in opens, you approve the same four scopes, and the five tools appear in your connector list.
Claude Desktop draws on the same connector configuration as the web app, so setting it up in one place covers both.
Two caveats worth knowing before you go looking. Custom connectors are not available on every plan, and if your account belongs to an organization, an administrator can restrict which connectors people are allowed to add. If the option is missing entirely, that is usually the reason rather than anything on FaithGPT's side.
Cowork
Cowork uses Claude's connector system, so the same URL works. I am not going to invent the click path. Anthropic has been shipping Cowork quickly and the navigation has moved, so open its settings, find where connectors are managed, and paste the server URL there. The OAuth flow behaves identically once you do.
Cowork is where this connector earns the most, and the reason is the shape of the work. Cowork sessions run longer and lean on documents in a way coding sessions do not. Preparing a study, reviewing a manuscript. That is exactly the work where a wrong verse survives all the way to the audience.
What to actually do with it
Sermon prep on the side of a build
You are shipping a church website on Thursday and the pastor needs the passage list for Sunday. Rather than switching context, you ask in the same session for passages on perseverance across both Testaments, get five back with references, and paste them into the notes file already open in the editor.
Doctrine review on generated copy
This is the workflow I would set up first. Marketing copy for a Christian product is very easy to get slightly wrong, and the wrongness is theological rather than grammatical, so a normal review misses it. Draft the about page, then hand it to review_doctrine before it goes near a designer. It flags where Scripture is doing work it was not written to do. Treat the output as a prompt to think rather than a verdict.
Devotionals from a shared repo
Put the server in .mcp.json, commit it, and your content team generates devotionals inside the same repository where the site lives, under the same account controls, with every draft anchored to a real passage. The tool publishes nothing, so a human still decides what ships.
Reference checking on anything you inherited
Point Claude at an existing content directory and have it verify every Scripture reference against lookup_scripture. This is dull work that nobody does, which is why old church sites are full of references that drifted years ago. A model grinding through it in a loop will find the ones a person skims past.
Limits, cost, and what will go wrong
Tool calls draw on your FaithGPT credits and are attributed to whichever account you signed in with. That is worth saying plainly: this is not a free side channel into FaithGPT, it is your account reached from a different client. If you burn through a batch of devotionals in Cowork, the credits come off the same balance the web app uses. Plans and credit amounts live on the pricing page .
Your client may ask you to approve each tool call rather than running them silently. That is deliberate. FaithGPT marks these tools as having side effects, because they write usage records and can spend credits, and a client that is honest about side effects will check with you.
The Scripture lookup covers KJV, ASV, and YLT. If you need a translation outside that set, this tool will not give it to you, and you should not let a model fill the gap from memory. That is the exact behavior the connector exists to prevent.
If the tools never show up, work through it in order. Run claude mcp get faithgpt and read the status. If it is not Connected, the OAuth flow probably did not finish, so trigger a tool call and watch for the browser tab. If nothing opens, check whether a popup blocker ate it. On claude.ai, confirm your plan allows custom connectors and that an admin has not restricted them.
You can also check the server itself without any client involved:
That returns FaithGPT's protected-resource metadata. If it responds and your client still cannot connect, the problem is on the client side.
When you want the API instead
MCP suits interactive work, where a person is in the loop and Claude decides which tool fits. For anything scripted, a nightly job or a server that needs FaithGPT answers with nobody watching, use the REST API instead. It is OpenAI compatible:
You authenticate with a key from the API dashboard rather than OAuth, and the full reference is at faithgpt.io/api .
The dividing line is simple. If a person is watching, use the connector. If a cron job is, use the key.
Start with one verse
Add the server, ask Claude for a passage you know cold, and check the output against your own Bible. Once you have watched it come back right, from a database rather than from a model's recollection, the reason to leave it connected becomes obvious.
You can see everything FaithGPT does with that same Scripture grounding at faithgpt.io .
Helpful Scripture resources
Contact & support
Email the FaithGPT team at hello@faithgpt.io or visit the contact page .
Explore FaithGPT
