Excel LLM Add-in

Call LLMs straight from Excel cells — and let an agent edit your sheet — across Excel for Mac, Windows, Web, and iPad. Bring your own provider (OpenAI, Mistral, Nebius, Scaleway, OpenRouter, or local Ollama).

▶ Live: https://crispstrobe.github.io/ExcelLLMAddin/ — install by sideloading manifest.xml (see install steps).

There are two editions in this repo:

Edition	Where	Best for
Office.js add-in (officejs/)	Mac · Windows · Web · iPad	store/AppSource, the widest reach, streaming
VBA .xlam (repo root .bas)	Mac · Windows	fully offline / air-gapped, single-file install

Both have the function set and the sheet-editing agent. The .xlam runs with no hosting and no internet (pair it with local Ollama); the Office.js edition adds Web/iPad, AppSource, and STREAM.

What it does

Worksheet functions (namespace LLM):

Function	Purpose
=LLM.PROMPT(text, [provider], [model])	Ask an LLM from a cell
=LLM.STREAM(text, …)	Like PROMPT, streamed into the cell live
=LLM.CLASSIFY(text, categories)	Pick one label from a range/list
=LLM.EXTRACT(text, instruction)	Pull a single value out of text
=LLM.FIELDS(text, fields)	Extract many fields into a row (text → table)
=LLM.TRANSLATE(text, language)	Translate a cell
=LLM.SUMMARIZE(text, [maxWords])	Summarize a cell
=LLM.SENTIMENT(text)	Positive / Neutral / Negative
=LLM.TAG(text, categories)	Multi-label: all labels that apply
=LLM.LIST(prompt, [count])	Generate a list, spilled down a column
=LLM.TABLE(prompt)	Generate a 2D table, spilled as a grid
=LLM.FILL(examples, inputs)	Infer a pattern from example pairs, apply down a column
=LLM.EDIT(text, [instruction])	Rewrite / fix grammar
=LLM.FORMULA(description)	Write an Excel formula from plain English
=LLM.EXPLAIN(formula)	Explain a formula (pair with FORMULATEXT)
=LLM.VISION(image, [question])	Ask about an image (URL or data URI)
=LLM.IMAGE(prompt, [w], [h])	Generate an image; wrap in =IMAGE(...) to show it
=LLM.ASK(question, context)	Answer a question using a range as context
=LLM.SIMILARITY(a, b, [model])	Semantic similarity (0–1) via embeddings
=LLM.RECALL(query, range, [k])	Semantic search: top rows in a range by meaning
=LLM.MAP(range, instruction)	Apply an instruction to every cell (batched)
=LLM.LIST_MODELS([provider]) · =LLM.CONFIG()	List models · show config

Plus an Agent: describe a change in plain English in the task pane and the model reads/edits your workbook (ranges, formulas, formatting, sheets) via tool-calling — with approve-before-apply by default. Its toolset can be extended with remote MCP servers.

Quick start

Use it (any user): sideload the hosted manifest — see officejs/PUBLISHING.md for Mac / org-deployment / AppSource.

Develop it:

cd officejs
npm install
npm start        # builds, starts the https dev-server, sideloads into Excel
npm test         # 160+ unit/functional tests (~99% line coverage), no Excel needed

Details, the browser dev-harness, and the CORS/proxy notes are in officejs/README.md.

Recipes

Drop these into cells (Office.js namespace shown; the .xlam uses the same names without the LLM. prefix, e.g. =CLASSIFY(...)):

Goal	Formula
Tag each support ticket	=LLM.CLASSIFY(A2, {"Bug";"Billing";"Feature"})
Split freeform text into columns	=LLM.FIELDS(A2, "name, email, company") (spills across)
Translate a whole column at once	=LLM.MAP(A2:A200, "translate to German") (batched)
Summarize to ~20 words	=LLM.SUMMARIZE(A2, 20)
Score sentiment	=LLM.SENTIMENT(A2)
Semantic match against a reference	=LLM.SIMILARITY(A2, $B$1)
Answer from a range of context	=LLM.ASK("what's the refund policy?", D2:D50)
Generate a list down a column	=LLM.LIST("EU capital cities", 10)
Generate a whole table	=LLM.TABLE("5 largest EU countries with population")
Fill by example (learn from 2-3 rows)	=LLM.FILL(A2:B4, C2:C100)
Write a formula from words	=LLM.FORMULA("sum col B where A > 100")
Explain a formula	=LLM.EXPLAIN(FORMULATEXT(D2))
Ask about an image	=LLM.VISION("https://…/chart.png", "what's the trend?")
Generate an image into a cell	=IMAGE(LLM.IMAGE("a red bicycle, studio photo"))

Agent (task pane, plain English): "In D1 put the sum of B2:B10, then bold anything over 100" · "add a column classifying each row of my selection as high/low". Writes are queued for one-click Apply unless you opt into auto-apply.

Providers

OpenAI · Mistral · Nebius · Scaleway · OpenRouter · Groq · Together · Cerebras · Google Gemini · Cohere · Hugging Face · Requesty · Ollama (local). OpenRouter, Nebius, and local Ollama work directly from the browser; others use the optional key-custody proxy (officejs/proxy/worker.js).

Both editions read the catalog from one source of truth, shared/providers.json; the Office.js tables are generated from it (npm run gen:providers) and CI fails on drift, so a provider URL can't go stale in one edition only.

Offline VBA edition (.xlam)

The VBA add-in (repo root .bas/.cls) is fully self-contained and offline — no hosting, no web server. It has near-parity with the Office.js edition:

• Functions: =PROMPT, =CLASSIFY, =EXTRACT, =TRANSLATE, =SUMMARIZE, =SENTIMENT, =ASK, =LIST, =FIELDS, =MAP, =SIMILARITY, =RECALL, =TAG, =EDIT, =FORMULA, =EXPLAIN, =LLMTABLE, =FILL, =VISION, =IMAGE_GEN, =LIST_MODELS, =LLM_CONFIG (modLLMFunctions.bas, modTasks.bas). (LLMTABLE is the Office.js TABLE, renamed to avoid Excel's legacy Data-Table function.)
• Agent (modAgent.bas) — run the RunAgent macro; the model edits the sheet via native Range tools (read, write, formulas, formatting, sheets, charts) with approve-before-apply. With local Ollama, this is a fully air-gapped AI that edits your workbook.
• MCP (modMcp.bas) — run SetMcpServer to add a remote MCP server's tools to the agent (best-effort JSON-RPC over HTTP; targets stateless servers).
• Providers: the same catalog as the Office.js edition — OpenAI, Mistral, Nebius, Scaleway, OpenRouter, Groq, Together, Cerebras, Gemini, Cohere, Hugging Face, Requesty, and local Ollama. Configure the added cloud ones via the settings menu's C option (they share one OpenAI-compatible code path).
• Solid plumbing: injected IHttpClient (WinHTTP/curl), real JSON (vendored VBA-JSON), UTF-8, a response cache, batched embeddings for =RECALL (one request for all candidates instead of one per row), and a RunAllTests self-test harness.
• Image generation: =IMAGE_GEN("...") returns a URL (pair with =IMAGE()), or run the InsertAIImage macro to drop the picture at the active cell. Set the key once with SetImageKey. (Black Forest Labs / FLUX; x-key auth.)
• Office.js-only: STREAM (VBA UDFs are synchronous -- no live cell updates).

Build: pwsh tools/Build-Addin.ps1 on Windows+Excel, or import the modules in the VBA editor and Save As .xlam (Excel is required to compile VBA — it can't be built on Mac or GitHub-hosted CI). Install offline: installer/mac-xlam/ builds a .pkg that drops the .xlam into Excel's add-ins folder. See installer/README.md.

License

AGPL-3.0.