tableski

tableski

Every spreadsheet is a table — tableski serves them to your agents in SQL.

Excel, CSV, Parquet, JSON — one penguin-family binary that turns your data files into queryable tables and serves them to AI agents over the Model Context Protocol.

• Single binary, stateless Streamable HTTP — built on the emperor-mcp framework (Emperor Profile P1 conformant: framed output, forwarded credentials, no session affinity).
• SQL engine powered by Apache DataFusion.
• Hatched in the kowalski rookery — the -ski is the family name.

Install

Prebuilt binaries (macOS arm64/x86_64, Linux x86_64, Windows x86_64) are attached to GitHub releases:

curl -fsSL https://raw.githubusercontent.com/yarenty/tableski/main/install.sh | sh

Windows: download the .zip from the releases page. Or via cargo:

cargo install tableski

Connect an MCP client

tableski is a standard MCP server over stateless Streamable HTTP — point any client at the server URL:

# Claude Code

claude mcp add --transport http tableski http://127.0.0.1:8080/

// Claude Desktop (claude_desktop_config.json) via mcp-remote

{ "mcpServers": { "tableski": { "command": "npx", "args": ["mcp-remote", "http://127.0.0.1:8080/"] } } }

Or raw HTTP: POST JSON-RPC to / with Accept: application/json (see below).

Run

# One model for every format: --file registers by extension, table name = file stem

(.csv .parquet .json/.ndjson/.jsonl; workbooks .xlsx/.xls/.ods = one table per sheet)

cargo run -- --file fixtures/sample.xlsx --file fixtures/sample.parquet \ --file fixtures/sample.ndjson --export-dir ./exports

(--csv/--xlsx flags still work; name collisions get _2/_3 suffixes)

Ask it something (tools: list_tables, query_sql, get_schema, column_statistics, export_result — all output framed as data per Emperor Profile E8). SQL joins work across formats: an xlsx sheet against a Parquet file against an NDJSON log. export_result writes a query's rows to .csv/.xlsx — only inside the --export-dir sandbox (relative names, no ..; the tool is disabled unless the operator passes the flag):

curl -s -X POST http://127.0.0.1:8080/ \

-H 'Content-Type: application/json' -H 'Accept: application/json' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"query_sql", "arguments":{"sql":"SELECT p.name, SUM(o.amount) AS total FROM people p JOIN orders o ON p.name = o.name GROUP BY p.name ORDER BY total DESC"}}}'

=> ada 150.5, linus 99.99 — a SQL join across two sheets of one workbook

Sheet names are slugified into table names (People → people, 2024 Orders → t_2024_orders); first rows become column names when they look like headers (--headers auto|first-row|none to override). Column types are inferred: integers, floats, booleans, dates (as timestamps), text.

Real-world workbooks are messy — tableski handles the mess: both Excel date systems (1900 and 1904), formula cells (cached values ingest; #DIV/0!-style errors become NULL), merged cells (top-left value, NULL continuations), display formats (raw values, always), ragged rows, format-only padding (trimmed), unicode sheet names, and a --max-rows cap that fails loudly instead of cutting silently. The nasty-workbook corpus lives in fixtures/corpus/ — every trait has a fixture and a test.

New here? The five-minute demo uploads a two-sheet workbook to tableski.io or a local tableski, asks the first question, and shows how to ask the rest in plain words with kowalski, Claude or Cursor. Try the whole local flow in one go: ./scripts/demo.sh (spreadsheet → question → SQL answer → exported workbook). demo.tape renders it as a GIF with vhs.

With kowalski, the questions come in plain words and the answers come back as a workbook. Its Spreadsheet analyst profiles the tables through tableski, the model writes one SQL query per question, and every number comes from the query engine:

Status: CSV, Excel (xlsx/xls/ods, hardened against real-world workbooks), Parquet, and

NDJSON all serve today; results export to csv/xlsx; binaries ship per release. Launch is next.

Serving clients you do not trust

By default tableski runs whatever SQL DataFusion accepts, which is right for a local single-user setup. Add --untrusted-sql when the server is reachable by agents or people you do not control:

tableski --file sales.xlsx --untrusted-sql

Then only read-only queries run: SELECT, WITH, VALUES, EXPLAIN, one statement per request, over registered tables only. CREATE, INSERT, UPDATE, DELETE, DROP, COPY, SET, CREATE EXTERNAL TABLE, table functions (read_csv(...), range(...)) and references to unregistered tables or file paths are rejected on the parsed statement, before anything executes, with an error that names the reason. DataFusion's own DDL/DML/statement switches are turned off as a second fence.

Limits

One query cannot take the machine down. Defaults, each 0 = unlimited:

Cartesian products (CROSS JOIN, or a join without an = condition) are the one shape a timeout cannot stop once running, so --untrusted-sql refuses them on the plan before execution.

The timeout covers planning as well as execution, so a statement that is expensive merely to optimise is cut too. What the pool does not meter: memory allocated by scalar functions (repeat, lpad, ...). Run a server for untrusted clients under a container memory limit, as the tableski.io deployment does. The whole set is exercised by tests/hostile_sql.rs (DDL/DML, filesystem reach, identifier tricks, blow-ups, parser stress, export traversal); add a line there when you find a new bad shape.

Use as a library

The engine is its own crate, tableski-core: register files into a DataFusion SessionContext, then run the tools with call_tool or serve AppState through any emperor-mcp transport. No HTTP, no CLI.

[dependencies]

tableski-core = "0.1"

let sources: Vec<Box<dyn TableSource>> = vec![Box::new(FileSource::new().path("orders.xlsx"))];

let tables = register_sources(&ctx, &sources, &IngestOptions::default()).await?; let state = AppState::new(Arc::new(ctx), tables); let grid = call_tool(&state, "query_sql", json!({ "sql": "SELECT count(*) FROM orders" })).await?;

Tables come from TableSource implementations. FileSource is the files-on-disk one; a source that streams or refreshes (a price feed, a file that is rewritten) implements the same trait and keeps its state behind its own DataFusion TableProvider, so every query sees the rows that exist when it is planned. The module docs of tableski_core::source spell out the contract; crates/tableski-core/tests/table_source.rs is a worked appending source.

For a server with tenants, Quota and UsageMeter (queries per day, bytes stored, files, rows per file) plug into the same state with with_quota; the meter is in memory, persisting usage is the host's job.

How it compares

Honest positioning — different tools solve different problems:

If you need to *edit* workbooks (styles, charts, cell writes), pair tableski with a writer-oriented server; tableski is the analysis and export engine.

Contributing

See CONTRIBUTING.md: run the checks, keep the docs in step, and sign off your commits (DCO, git commit -s).

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)

at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in tableski by you shall be dual licensed as above, without any additional terms or conditions.