Landmark

An MCP server that lets you ask a spreadsheet questions out loud, instead of walking it cell by cell.

Built for the Alexa+ track of the Amazon Developer Hackathon 2026. MIT licensed. A self-hosted MCP server on protocol revision 2025-11-25 over Streamable HTTP, with a browser voice client that stands in for a device.

The problem, stated accurately

It is tempting to write "screen readers lose header context." That is not true, and getting it wrong would be the fastest way to lose the trust of anyone who actually uses one. JAWS and NVDA both announce headers in Excel.

The real difficulty is narrower and harder:

• Header association is manual, per worksheet, and does not travel. Headers set up for one reader do not carry to another. NVDA #11801, open since November 2020, is about exactly that: headings defined with NVDA's commands are not spoken in JAWS, although NVDA's user guide says they will be.

• Authors rarely declare a table's structure, even where they can. The WebAIM Million, February 2026 counted 948,225 tables on the top million home pages and found valid data-table markup on 19% of them. That is web pages, not spreadsheets, but it is the largest count we know of how often people mark up headers when a way to do so exists. For spreadsheets, Section508.gov puts it plainly: "Excel does not provide tools to make complex tables accessible."

• Even with headers announced, questions cost traversal. "What is the total?" is O(rows). "Which region did best?" is O(rows × cols). There is no cheap way to learn a table's shape before committing to walking it.

A sighted reader answers all three with a glance, which costs nothing and is not sequential. Landmark reconstructs the structure the file never declared, so the same questions can be asked rather than traversed.

Related MCP server: mcp-excel

What it does

Word for word, from the bundled budget file (npm run demo prints these and more, from the real server):

You:       what's in the budget file
Landmark:  "FY2026 Departmental Budget" has 5 rows and 3 columns. The columns are
           Department, Line item and Amount. 3 cells take their label from a merged
           block, so they are not the blanks they look like.

You:       total amount for engineering
Landmark:  560 thousand. That is the total of Amount across 3 rows.

You:       how do you know
Landmark:  That came from C3 through C5 on Budget. Each one is Amount.

The third exchange is the one that matters. At CHI 2026, Perera, Ananthanarayan, Goncu and Marriott studied how 12 blind spreadsheet users check what generative AI gives them ("I'm Always a Little Skeptical of It": Verification Practices of Blind Users When Working with Generative AI in Spreadsheets). Their participants never fully trusted the output without verifying it, and verifying was often effortful or time-consuming. Over audio there is no cell to glance at. So the arithmetic happens in the server, not in a language model, every answer keeps the cells it came from, and table_explain reads them back.

For a highest or a lowest, it names the row, and the explanation starts with the winning cell:

You:       who is the top rep                                (the sales table)
Landmark:  21 thousand, for Chi. That is the highest Revenue across 5 rows.

You:       how do you know
Landmark:  That came from C4 on Sales, for Chi, the highest of 5 cells: C2 through C6.
           Each one is Revenue.

Cells it could not count are said out loud, because "the total" and "the total of the rows that held numbers" are different claims. A table_query on a five-row CSV with one blank cell and one "n/a" (made for this check, not in the bundled data):

Landmark:  650. That is the total of Sales across 3 rows. I skipped 2 rows: 1 was empty
           and 1 did not hold a number.

You:       how do you know
Landmark:  That came from B2, B4 and B6 on gaps. Each one is Sales. 2 cells did not
           count: B3 empty and B5 not a number ("n/a").

Try it on the bundled tables

data/index.json holds six small tables, named after their fixture files: five workbooks generated by test/fixtures/make.ts (sales, a stacked header, a merged header, the budget, and one sheet holding three tables) and 06-countries.csv, eight countries with population, GDP per capita and region. That CSV's own note calls its figures "illustrative figures for testing only"; they appear to match World Bank World Development Indicators estimates (CC BY 4.0), so treat them as public data, not ours.

Run the server (below), open http://localhost:8787/ in Chrome or Edge, and either press Ask by voice (or Space), wait for the tone and speak, or type in the Type a question box and press Ask. Esc interrupts. Turn Speak answers off to have answers announced through your own screen reader instead of the page's voice.

Questions like these are answered; each was checked against the running server:

• what's in the budget file, total amount for design, break it down, which department spent the least

• open the countries table, compare the population of Vietnam and Thailand, what regions are there, total population not in Asia, which countries have a population over 100 million, what region is Peru in

• open the sales table, who is the top rep, rank the reps by revenue, total revenue excluding the east, how many deals closed in August, how many deals closed after July 4, which region has the most reps

• open the three regions file, how many tables are in this file, table 2, total actual

• check the structure, use the first row as headings, the top two rows are headings, row one is data

• how do you know, save my place, carry on, more, help, start over

It cannot work out a median or percentiles, a difference per group ("which region grew the most"), or synonyms it was not told ("richest", "populous", "people"). Those get a question back or a plain refusal, never a guessed number:

You:       what is the median population
Landmark:  I cannot work out a median. I can give a total, an average, the highest,
           the lowest, or a count.

You:       which country is the richest
Landmark:  I could not find "richest" in this table. The columns are Country,
           Population, GDP per capita (usd) and Region.

Why this is not a basic MCP wrapper

The Alexa+ track rules name "a basic MCP wrapper around an existing API" as the obvious shape of entry, against a creative list that includes state across sessions, MCP Apps and Agent Skills. That is a fair challenge to put to a project shaped like this one, so here is the answer in specifics.

1. It infers structure that is not in the file. A spreadsheet declares a grid, not a table. Landmark detects table regions separated by blank rows, scores which rows are headings and says when it is unsure, resolves merged blocks, and rebuilds a full heading path for every column. Four columns all labelled Revenue come back as 2026 Q1 Revenue … 2025 Q2 Revenue, with the path ["2026", "Q1", "Revenue"] kept in the structured result. A row labelled Total, Subtotal or Tổng cộng is recognised and left out of answers, and the answer says so ("300. That is the total of Amount across 2 rows. I left out the Total row."). So is a "Housing Total" row of the kind Excel's Subtotal command and PivotTables write, or a hand-typed "Total Expenses" or "Tổng doanh thu", once its figures add up the rows above it. Numbers written the Vietnamese or European way ("1.234,56", "45.000 ₫") are read as such; a value that could go either way ("12.500") is read with the dot as the decimal point, and the description and every total, average, highest, lowest or breakdown of that column say so out loud. Dates such as 01/09/2026 are read day first when the rest of the file is written that way; when nothing in the file settles it, the reading is stated as a guess. When the reading of the headings is wrong, the listener can correct it by voice:

You:       check the structure                     (table 3 of the three regions file)
Landmark:  I am treating 1 row as headings — the first row reads as labels. That gives
           the columns Note and Status. I am not certain: I could instead treat every
           row as data. Tell me how many rows are headings to change it.

You:       row one is data
Landmark:  Right — I now read no heading rows instead of 1 heading row. That gives 2
           rows of data, and I will call the columns by position. That holds for the
           rest of this conversation.

There is no API to wrap here; this is the work.

2. Provenance is a tool, not a log line. Every aggregate records its source cells, the rows it left out, and why. Nothing is recomputed on explain, because a recomputation that disagreed with what was already spoken would be worse than no provenance at all.

3. Output is shaped for listening in the server, not left to a prompt. Every tool returns a spoken field: a ready-to-say summary capped at about 70 words (roughly 30 seconds of speech), at most five items at a time by default with a cursor for the rest, numbers scaled for listening ("about 100.4 million"), and no tool names, ids or field names. The exact figures stay in the structured result. The server's instructions offer spoken as ready to say and ask the host to say it as written. That is an offer, not a script: Amazon's add-on design guidance says you influence Alexa's response through the data you return, not by scripting it directly, so a host that writes its own reply still has the exact numbers and cells to write it from. Errors are spoken too, with a next step, including arguments the schema refuses and tools that do not exist.

4. State outlasts the request, and where it lives is stated. Answers, heading corrections and bookmarks are kept per conversation. On Cloudflare Workers each conversation is a SQLite-backed Durable Object, cleared after 90 days without use, and each answer is its own Durable Object, deleted after 7 days. Bookmarks carry across conversations only for the voice client, which keeps one random id per browser in localStorage; for MCP hosts they, and structure corrections, last one connection. Answers can be explained from any later connection by their id, for a week on the Worker. Run locally, all of this is in memory and ends when the server stops. Callers that send no session id share one in-memory conversation, and GET /health says which of these you are getting ("state": "durable" or "this instance only").

5. It works across sheets and files. table_compare runs two aggregations over different columns, sheets or files and returns the difference and its direction in one sentence ("2026 Q1 Revenue is 200 more than 2025 Q1 Revenue, about 11%: 2100 against 1900."), so nobody has to hold two numbers in their head and subtract them. It takes the same conditions as a query and says them first ("For South, Target is 5100 more than Actual, about 131%: 9000 against 3900.").

6. An MCP App shows the cells. table_explain declares a UI resource, ui://landmark/explain (text/html;profile=mcp-app), that draws the source table with every counted cell highlighted and marked "(counted)" in text as well as colour, and the winning cell of a highest or lowest marked "(the answer)", up to 40 rows. Many blind and low-vision people have some usable sight or work beside sighted colleagues; this lets them, or the colleague, see the cells the sentence names. It never replaces the spoken answer. Its handshake and rendering are tested in Node against the MCP Apps message shapes (test/engine-widget.test.ts); it has not yet been rendered in a real MCP Apps host, and the bundled voice client does not show it.

7. An Agent Skill teaches a model how to use it. skills/landmark-tables/SKILL.md is an Agent Skill for an assistant driving these tools: describe before querying, aggregate instead of reading rows, say the spoken sentence, offer the working after a total, carry the cursor on "more". It has not been evaluated with a model yet.

And what it deliberately does not do: there is no get_cell, get_row or get_column. Those would rebuild the cell-by-cell maze in tool form and hand the traversal problem back to the model.

Prior art, conceded

The intersection is unoccupied; none of the neighbourhoods are.

The honest version of the market claim. We know of no MCP server that lets a blind user interrogate a spreadsheet by voice, and that is the gap this fills, but it is an observation, not a claim of being first. Two things cut against a stronger version. The official registry does contain end-user assistive servers: NeuroDock publishes several for cognitive and executive-function support, aimed at neurodivergent users rather than at developers. And the registry's search matches substrings of server names only, not descriptions or audiences, so a query returning nothing for "assistive" is evidence of nothing. An earlier draft of this README said the registry contained no assistive server whose consumer is a disabled end user. That was wrong, and it was wrong in the direction that flattered us.

Run it

Needs Node.js 22.7 or later and npm. Every script runs the TypeScript directly with --experimental-strip-types; 22.6, the first release to have it, cannot parse this code. Built on Node 23.11 on Windows 11, and tested on every push on Linux, macOS and Windows with Node 22 and 24, plus 22.7 on Linux. These commands are the same in bash and PowerShell:

npm install
npm run serve

Then open http://localhost:8787/ for the voice client. The MCP endpoint is http://localhost:8787/mcp (POST only) and GET /health reports the protocol revision, the number of tables and rows, and where state is kept. The bundled index ships in data/index.json, so there is nothing to generate first.

• Another port: npm run serve -- 8788 works in both shells, as does setting PORT (PORT=8788 npm run serve in bash, $env:PORT=8788; npm run serve in PowerShell).

• npm start builds with tsc first, then runs the compiled dist/src/local.js. npm run dev restarts on changes.

• The server listens on 127.0.0.1 only and refuses requests whose Host header is not a loopback name. HOST=0.0.0.0 opens it to the network, with no Host check. Behind a tunnel, name the tunnel's host in LANDMARK_ALLOWED_HOSTS instead; that also admits its https:// origin for the voice client.

npm run serve, npm run ingest and npm run demo run the TypeScript directly, so Node prints an ExperimentalWarning about type stripping first; it is harmless.

Your own spreadsheet

Files are loaded by an offline ingest step, not uploaded at runtime. It reads .xlsx, .xlsm, .csv and .tsv, prints what it inferred for every table, and writes an index. This form works in both bash and PowerShell (PowerShell swallows a bare --, so it is quoted):

npm run ingest '--' path/to/budget.xlsx path/to/*.csv --out data/mine.json

Then start the server on that index:

LANDMARK_INDEX=data/mine.json npm run serve                  # bash
$env:LANDMARK_INDEX="data/mine.json"; npm run serve          # PowerShell

The CLI expands wildcards itself, so patterns work in any shell. If any input cannot be read, nothing is written. --out defaults to data/index.json, the demo's index; writing elsewhere keeps the demo intact (data/* other than index.json is git-ignored). Running npm run ingest '--' test/fixtures/*.xlsx test/fixtures/*.csv rebuilds the bundled index. Apart from its timestamps, the committed copy differs from a fresh build only in the heading-evidence scores recorded for each table, which no tool reports, so every answer comes out the same.

Deploy to Cloudflare (optional)

The hackathon FAQ says hosting is not required: a locally runnable public repository and the demo video are enough. To put it on the internet anyway, on the Workers free plan:

1. Create a Cloudflare account.

2. npx wrangler login

3. npx wrangler deploy. If the account has no workers.dev subdomain yet, Wrangler asks you to register one.

4. Open https://landmark-mcp.<your-subdomain>.workers.dev/health and check it says "state": "durable".

The two Durable Object classes are declared in wrangler.jsonc and created by the first deploy; there are no ids to paste and no KV namespace. npx wrangler deploy --dry-run shows both bindings (LANDMARK_STATE and LANDMARK_ANSWERS) and a bundle of about 970 KiB, 195 KiB gzipped. The voice client is served from the same origin as /mcp. Other browser origins get 403 unless listed in the LANDMARK_ALLOWED_ORIGINS variable; no CORS headers are sent. To run the Workers build locally with its Durable Objects, use npx wrangler dev.

This repository has not been deployed by its author yet, so there is no public URL here.

How it is put together

spreadsheet ──[ ingest CLI, Node ]──▶ data/index.json ──[ server ]──▶ MCP tools
                                                            │
                                   voice client (web/) ─────┘  same origin, POST /mcp

The split is deliberate. ExcelJS wants Node streams and Buffer, which Workers provide only behind the nodejs_compat flag, so parsing there is possible but not free. A spreadsheet does not change between deploys, so everything expensive happens once, offline, and the server does pure computation over plain JSON. The same fetch handler runs under Node locally and on Workers.

Measured on the Windows 11 machine this was built on: a grouped sum over 2,000 rows, through the full handler in one process, took 2.7–6.5 ms at the median and 7.1–19.3 ms at the 95th percentile over six runs of test/latency.test.ts, which asserts that the 95th percentile stays under 500 ms. Network time is on top of that. Start-up on Cloudflare itself has not been measured.

The nine tools

Tests

npm test            # 439 tests
npm run typecheck   # strict, noUncheckedIndexedAccess, exactOptionalPropertyTypes
npm run demo        # prints what the server says to a scripted conversation

test/demo.test.ts pins the filmed demo conversation word for word, and test/judge-questions.test.ts pins questions a judge might make up. GitHub Actions (.github/workflows/test.yml) runs the typecheck, the suite and the build on Linux, macOS and Windows with Node 22 and 24, and on Node 22.7, the oldest supported. The suite also passes with the process time zone set to America/Los_Angeles, Pacific/Kiritimati and Asia/Ho_Chi_Minh. In Git Bash on the Windows machine those runs were made on, a TZ set on the command line did not reach Node (in PowerShell it did), so the runs used a small preload script that sets process.env.TZ before the tests start.

Status and honest limits

• No blind or low-vision person has used it yet. The CHI 2026 study above worked with 12 blind spreadsheet users; this has been tried by none, and that gap is real.

• You cannot bring a file at runtime. A spreadsheet reaches Landmark only through the offline ingest CLI, followed by a restart of the local server (or LANDMARK_INDEX pointing at the new index) or a redeploy of the Worker. A deployed Worker serves the index it was built with to everyone.

• English only. Speech recognition and synthesis are set to US English and the voice client's routing rules are English. Column names in other languages are kept as written, but an English voice reads them.

• Speech input depends on the browser. It uses the Web Speech API's recognition, which Chrome and Edge provide; Safari also exposes it but has not been tested here, and Firefox does not, so there you type. Chrome's recognition is server-based, so the audio of a question leaves the machine even when Landmark runs locally. The typed box avoids that. The microphone path and screen readers (NVDA, JAWS, VoiceOver) have not been tried on real hardware with this build; the typed path has been checked in a browser.

• The voice client is a rule router, not a language model. It handles the phrasings above and asks when it is unsure. It was probed with several hundred improvised questions, each checked against figures worked out separately from the rows, and every wrong answer those probes found is now a regression test — but a phrasing outside its rules can still be misread rather than refused. "How do you know" names the exact rows and cells behind any answer, which is the way to catch that. A model-driven MCP host would not be bound to those rules, but none has been tested.

• Heading detection is a heuristic, and says so. When it is unsure it tells the listener and offers the alternative, and the listener can correct it by voice.

• Ambiguous numbers are a guess. A column whose only values look like "12.500" is read with the dot as the decimal point, and the description and every total, average, highest, lowest or breakdown of that column warn that the figures may be a thousand times larger.

• Not tested on Alexa+. Amazon's Alexa+ MCP Toolkit is available to select partners only; the hackathon FAQ says there is currently no way for hackathon participants to apply for or gain access. The MCP server is the deliverable the track asks for, and the browser client stands in for a device. How an Alexa+ host holds MCP sessions, and therefore how long corrections and bookmarks last there, is unknown.

• No authentication. A session id separates conversations; it is not a password. Anyone who can reach a deployed endpoint can read every bundled table. Text from cells, headings and notes is passed to a host as data, and the server's instructions say it is never to be followed as instructions.

Licence

MIT; see LICENSE. Dependency audit notes are in docs/SECURITY-NOTES.md; the two moderate advisories npm audit reports are analysed rather than suppressed.