Build1 publisher2 min readPublished
An Excel Desktop sidebar routes every AI request through a validated action list
A developer's Copilot alternative for Excel Desktop sends natural language to OpenCode and gets structured actions back, which the add-in checks before running. It only runs the actions somebody wrote by hand.
The Engineer · Build desk
What happened
- A developer released an Excel Desktop add-in with a Copilot-style AI sidebar that uses OpenCode as its AI interface instead of Microsoft's Copilot experience.
- The model never touches the workbook directly: it emits structured actions, and the add-in validates them and executes them itself.
- Requests travel from the sidebar through Office.js to a local Python bridge called bridge_server.py, then on to OpenCode and the AI model provider.
- Documented coverage includes pivot tables, charts, conditional formatting, formatting, formula operations and structural changes.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Capability is bounded by whoever writes the action types, because a request with no matching action cannot fall back on improvised script, so the ceiling on this design is a standing maintenance commitment.
- cost An adopter pays in local software and provider tokens rather than a per-seat licence, and owns a Python process on every workstation that runs the sidebar.
- decision Teams whose objection to Copilot is the usage allowance now have a reference implementation to price against, and the live question becomes who staffs the action schema.
- contradiction The saving is asserted by the builder and not measured, so the case for replacing a licence currently rests on one user's editing pattern.
The published example action is a sort: `{"type": "sort_range", "sheet": "Sales", "range": "A1:F500", "column": 6, "ascending": false}` [5]. A validator in front of that can check that the Sales sheet exists, that A1:F500 parses, that column 6 falls inside it, and that `ascending` is a boolean. It cannot check that column 6 holds revenue. The model turned "Sort the sales data by revenue descending." [10] into an integer index before the add-in saw anything, and Office.js will sort column 6 either way.
The narrowing is deliberate. The post says the model "doesn't get unrestricted access to execute arbitrary JavaScript", and calls that intentional [6]. I think that is the right call for a tool pointed at a live workbook, and it sets the price: every operation you want has to exist as an action type, with a validator and an Office.js implementation behind it [4]. Past the documented set, a new capability means writing code.
Adoption means running more software than a sidebar. Three components sit between the sidebar and the model provider [1], one of them a Python bridge process on the workstation next to Excel [2]. The design lets you pick the AI layer separately. "Why should the Excel interface and the AI provider have to be the same product?" the author wrote [11].
The reason for building it was metering. The author reports that when he experimented with Microsoft Copilot for Excel, a relatively small number of interactions and edits could hit usage limitations [7]. He wrote that he wanted an AI assistant "that I could actually use for long Excel sessions without constantly worrying about the Copilot usage allowance" [8]. For that to be your reason too, your sessions have to look like his: a long run of small edits, each one its own request, on one workbook.
As a licence argument this is thin. The post gives no cost-per-user comparison against a Copilot seat and no figures on how often the generated actions were right [13]. It supplies a working separation of the three layers: Excel stays the spreadsheet application, the sidebar talks to an AI system through OpenCode, and JavaScript performs the operations [12]. Whether that beats a seat depends on your token spend and on who owns the action schema after the person who wrote it moves on.
What to watch
- Whether the repository publishes the full action schema and validator, which would show how much of Excel is actually reachable through it.
- An independent run of the same prompts against both the add-in and Copilot for Excel, with task success rates recorded.
- Whether Microsoft states concrete interaction limits for Copilot in Excel, so the metering complaint can be checked against a number.