MS Excel AI Plugin

The last AI plugin you install in Excel. Open source, bring-your-own-model, local-first.

Connect any AI model — Claude, OpenAI, Gemini, OpenRouter, any OpenAI-compatible endpoint, or fully local models via Ollama / LM Studio — and let it understand, analyze, and modify your workbooks as a real agent, with every change previewable and undoable.

Why another Excel AI tool?

Because every existing one locks you to a single vendor, charges a subscription, writes into your sheets with no preview, forgets everything between sessions, and can't see anything outside the open workbook. This project is built around exactly those gaps:

• 🔑 Bring your own model — your keys, your endpoints, your local models. Works 100% offline with Ollama.
• 📦 Optional built-in local model — flip one switch and the plugin works with no API key, no Ollama, no setup (a small Apache-2.0 model like Qwen3-4B, run in-process). Strictly opt-in: it downloads nothing and consumes zero memory or compute until you enable it, loads lazily on first use, and unloads itself when idle. A project fine-tune specialized for spreadsheet work (ExcelLM) is on the roadmap.
• 🛡️ Safe agentic edits — every AI change is a reviewable change-set: preview → approve → one-click undo, with a full audit log.
• 🎯 Works well with any model, weak or strong — a model-agnostic harness does the heavy lifting: capability probing grades each connected model, scaffolding adapts to it, every tool call is schema-validated, every edit is verified after the fact, and ready-made playbooks cover the jobs Excel is actually used for (reconciliation, cleaning, budgeting, reporting, merging) so precision comes from the system, not the model.
• 📊 Built for big workbooks — a workbook indexer and formula dependency graph let the agent work on 50k-row, 20-tab models without blowing the context window.
• ⚡ Bulk AI functions — =AI(), =AI_CLASSIFY(), =AI_EXTRACT() and friends, batched and cached, routed to cheap or local models.
• 🧠 Real memory — per-workbook and per-user memory that makes every session smarter than the last. Local, visible, editable, exportable.
• 🔌 Open data connectivity — import PDFs, Word docs, and scanned tables into your sheet with zero setup (bundled converter), read other local workbooks and CSVs, and connect databases, SaaS tools, and APIs through the Model Context Protocol if you want to go further.
• 🔒 Private by default — a local companion app keeps keys in your OS keychain and all data on your machine. No accounts, no telemetry, no server.

Architecture in one paragraph

A thin Office.js add-in (taskpane + custom functions) acts as the agent's "hands and eyes" inside Excel, and a local companion app (the sidecar) running on localhost hosts everything stateful and sensitive: the model provider layer, the agent loop, the memory store, the MCP client, the batch engine, and your API keys.

📐 Design document

The full design — problem analysis, competitive gaps, architecture, safety model, memory system, MCP integration, and the phased roadmap — lives in docs/DESIGN.md.

Status

🚧 Phase 1 (Foundation) implemented — see the roadmap:

• ✅ Monorepo: shared/ (tool schemas + WS protocol) · sidecar/ (Node daemon) · addin/ (Office.js)
• ✅ Model layer: Anthropic, OpenAI, and any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM, …)
• ✅ Agent loop with schema-validated tools, repair-retry for malformed calls, and emulated tool calling for models without native support
• ✅ Change-sets: stage → preview (with reasons) → approve → verify (auto-rollback on introduced errors) → one-click undo
• ✅ Sidecar: pairing-code auth, provider settings API, =AI() function batch engine with caching, cost ledger, FTS-backed memory store
• ✅ Headless workbook emulator with a real formula engine, so the entire stack is tested without Excel: 40 unit/integration tests + a full-stack E2E smoke test (pnpm smoke)
• ⬜ Workbook indexer, playbooks, built-in local model, MCP, document import — next phases

Run it

pnpm install        # Node >= 22.13
pnpm check && pnpm test && pnpm smoke   # typecheck, tests, full-stack E2E
pnpm build && pnpm dev:sidecar          # start the sidecar (prints a pairing code)

Then follow docs/INSTALL.md — sideload addin/manifest.xml into Excel, enter the pairing code in the task pane, pick your model (Ollama works free and fully local; Anthropic/OpenAI keys also supported), hit Test connection, and chat.

Current dev targets: Excel on Windows and Excel on the web (Chromium). Excel for Mac needs an HTTPS-localhost story (WKWebView blocks mixed content) — tracked in the design risks.

License

MIT © the MS-Excel-AI-plugin contributors.