AI chat & agents

AI in Lamp Bible is optional, Mac-only, and runs through an AI tool you already use. Lamp Bible has no AI service of its own: it starts your provider’s command-line tool on your Mac, and gives it read-only tools for looking things up in your library.

Turning it on

AI is off until you turn it on. Until then, Settings → AI & Agents shows one switch, Show AI Features, and nothing else in Lamp mentions AI. Off hides Reader Chat, the writing agent and plans made from a description, turns off the system’s Writing Tools in Lamp’s text, and withdraws what other apps may read through Lamp. Your chats and drafts are kept for when you turn it back on. A library that was already using AI starts with it on.

The switch is shared with your iPhone and iPad, where it is Settings → AI Features and covers making a reading plan from a description.

Setting up

Three providers are supported. Each needs its command-line tool installed and signed in:

With AI on, Settings → AI & Agents lists each provider under AI Provider Accounts, and whether its tool is installed and connected. Connect… starts the provider’s own sign-in. Lamp Bible never reads or stores your credentials; they stay with the tool. The tool must be on your login shell’s path, which is where installers normally put it.

What the AI may see

Lamp’s tools are read-only: the AI can look up passages, commentary, dictionary entries, books, plans, quizzes and — if you allow it — your own study, and analyse a Strong’s number, but cannot change your library. The two exceptions let another app replace a writing’s draft or add a reading plan, and are off unless you turn them on; see Using Lamp in other apps. Test Module Tools checks that the tools are working.

Provider, model and effort

At the top of each chat, three menus choose the provider, the model and the effort. Lamp names no model of its own: the menu offers the provider’s default, the names the provider itself keeps pointed at its current models — opus, sonnet and haiku for Claude Code — and Other Model… for a name you type, which is passed on as it is. Effort is low to max for Claude Code, a value you type for Codex, and for OpenCode the variants the installed tool lists for the model you named. The choices belong to the chat, so two tabs can use two providers or two models at once.

Study skills

Lamp’s agents come with two skills for questions about scripture, which they read before answering one:

• sola-scriptura tells what a passage states from what is inferred from it or brought to it. Asked about one claim, it says whether scripture establishes it, in a paragraph with the text quoted. Given a whole piece of writing, it goes through it claim by claim — stated, entailed, inferred, imported, unsupported or contrary — and ranks what it finds by how much of the argument rests on each.
• original-languages reads the Hebrew or Greek behind a verse, word by word with each word’s parsing, and shows where translations differ and whether the difference is a translator’s decision. It turns to TBESG first for Greek and to BDB for Hebrew, and follows a Greek word to the Hebrew it renders in the Septuagint. Given a writing, it checks the claims it makes about the original.

Both measure against the text alone. They look passages up rather than quoting from memory, report a commentary’s view as that commentary’s, and never count what your library happens to hold as evidence for a reading. Reader Chat and writing agents have both; other apps get them as prompts.

Reader Chat

Reader Chat (⇧⌘J, or the speech-bubble button) opens beside the reader once AI is on. Ask about the chapter on screen, or name any other passage. It answers from your modules through the read-only tools, and cannot edit files or run commands.

Each message is sent with the chapter and translation you are reading, the verse marked in the reader, and any words you have selected in it, traced to their verses. With Include personal writing, notes, and highlights on, it also carries your own notes and highlights in that chapter; with it off, the chat is told they are not shared. A line under the chapter’s name says what the next message will carry.

The chat looks only in the modules your Module Access setting allows, so a module you have hidden stays out of its answers unless you choose all installed modules. A line at the top of the chat says what it can reach.

Replies appear as they are written, and with Claude Code a line says what the agent is doing meanwhile — the tool it is using, and the passage it is reading. Afterwards, a line under your question names the tools its answer used. A reply keeps arriving if you close the chat, switch it for Study Tools or leave the reader, and is there when you come back. A message you send while a reply is still arriving waits its turn: until it goes, you can reword it, send it sooner than the others waiting, or remove it, and it is still waiting if you quit.

Each reader tab keeps its own chat, and comes back to it when you relaunch. Chats starts a new one or goes back to an earlier one, each listed by how it opened; its search field finds a chat by anything said in it, by you or in a reply. A chat open in another tab is marked as such. Right-click a question and choose Fork From Here to carry the conversation on from that point as a chat of its own, leaving the original exactly as it was. Keep chats for, in Settings → AI & Agents, sets how long a chat is kept after it was last used — two months unless you change it — and a chat nothing was said in is not kept at all. Reader Chat conversations stay on the Mac and do not sync.

Writing with an agent

In the writing editor, the Agent button opens a workspace for that piece of writing. The agent works on the same text you do: the draft in the editor and the agent’s copy are kept in step as either of you changes it, and every change goes into Revisions, where you can undo it or restore an earlier version. If you and the agent change an open file at once, you choose whose version to keep. The agent is told which file is open in the editor and the text you have selected in it, so you can point at a paragraph and ask about it.

• Chat is a conversation with the agent, headed by the workspace folder it reads and edits. Its edits go straight into the draft, and each one is in Revisions. When Claude Code needs permission — to run a command or fetch a web page, say — it asks in the chat: Allow, Always Allow for that exact kind of request, or Deny. A message sent while the agent is working waits its turn, as in Reader Chat, and closing the Agent pane does not stop it. Undo This Turn…, on a question’s right-click menu, puts the draft back as it was before that turn and takes the turn, and everything after it, out of the conversation. Anything you wrote since is replaced, but the restore is recorded in Revisions, so you can put it back.
• Terminal runs the provider’s own tool directly, continuing the same session, for anything the chat does not offer.
• Context holds files you add for the agent to read — sources, a previous talk, notes — as read-only copies.
• Workspace Skills are instructions the agent can follow. Three are built in: build-lamp-deck, which turns the writing into a Slide Studio deck, and the two study skills, which can audit the draft. You can write your own, and add one you wrote for another writing.

The workspace — companion files, context, skills, revisions and the conversation — syncs with your library, so it follows the writing to your other Macs.

Using Lamp in other apps

The same library tools can be used from an AI app you already run — Claude Desktop, Codex, Cursor or opencode — through the Model Context Protocol. Turn them on in Settings → AI & Agents, under Use Lamp in Other Apps, once AI is on:

All three need Allow agents to query Lamp modules, and the last two need reading allowed as well. What an app may read follows the same Module Access and personal-content settings as Lamp’s own agents.

Setting an app up

Once reading is allowed, the section lists the apps it can set up:

• Add to Claude Desktop… builds a bundle and opens it in Claude Desktop, which asks you before installing it.
• For opencode, Codex and Cursor — or Claude Desktop by hand — Copy puts a snippet on the clipboard, and the section names the file to paste it into.

Lamp never edits another app’s settings. Each snippet names a small launcher in Lamp’s Application Support folder, which finds the app wherever it is installed, so the setup keeps working when Lamp moves or updates.

What the app can do

An app set up this way has the read-only tools a writing agent has: it can search your library, read passages, commentary, dictionaries, books, plans and quizzes, analyse a Strong’s number, check, fill in and draw a deck, and lay out a reading plan without adding it. An app that offers prompts also gets Lamp’s skills as prompts: Build a Lamp deck, Audit writing against scripture alone and Read the Hebrew or Greek.

With Allow other apps to add reading plans on, it can also add a plan over the passages you name — so a model you already pay for can work out a plan with you, show you its days, and put it in your library. Lamp lays the plan out and adds it when it next reads its library, and includes it in Today. It adds plans and nothing else: it cannot change or remove anything already there.

With personal content allowed, it can also read the writings that have an agent workspace — any writing whose Agent pane you have opened. With Allow other apps to change my writings on, it can replace one’s draft:

• The change is refused if you or another agent have changed the draft since the app read it, so it cannot write over words it has not seen.
• Lamp takes the new draft in when you next open that writing’s agent workspace, and records it in Revisions like any agent’s edit. If you have edited the writing in the meantime, your text stays, and the other app’s version is kept in Revisions instead.

Privacy

When you use AI, your messages — and whatever the agent reads to answer them — go to the provider you chose, under that provider’s terms and privacy policy. Nothing goes to Lamp Bible, which does not operate or proxy these services and never sees your conversations. If you never connect a provider, no AI tool is ever started. What another app reads through Lamp’s tools goes to whichever AI provider that app uses, under its terms. See the app privacy declaration.