Crate tableski 

Stateless Streamable HTTP (JSON + SSE) MCP server over the tableski engine.

Everything about tables and tools lives in tableski_core and is re-exported here, so tableski::AppState and tableski_core::AppState are the same type. This crate adds the HTTP wiring (app_router) and the CLI (src/main.rs). No Mcp-Session-Id is issued or required — every POST is independent.

Modules

excel

Excel ingestion: calamine → Arrow RecordBatch → registered DataFusion tables.

export

Result export: run a query and write the result set as CSV or XLSX — only into the operator-configured export directory (no caller-controlled paths outside it).

guard

Untrusted-SQL guard: decide on the parsed statement whether a client that we do not trust may run it. Read-only queries over registered tables pass; everything else is rejected before DataFusion sees it.

limits

Per-query limits: a memory pool cap on the session, a wall-clock timeout, and caps on the rows and bytes a result may carry back. Unlimited by default for embedders; the CLI applies QueryLimits::safe unless told otherwise.

quota

Quota primitives: what a tenant may use, and an in-memory meter that enforces it.

register

Unified file registration: one model for every supported format.

source

Where tables come from: the TableSource seam between a supplier of tables and the DataFusion SessionContext that serves them.

Structs

AppState

DataFusion-backed MCP handler: registered tables plus the session context to query them.

Collected

What collect_limited brought back.

FileSource

Files on disk: the CLI’s --file (extension picks the reader, table = file stem), --csv with an explicit table name, and --xlsx (one table per sheet).

IngestOptions

Ingestion options for register_workbook.

QueryLimits

Bounds on what one query may consume and return. None means unlimited.

Quota

The allowance of one tenant. None = unlimited.

SheetInfo

One registered sheet: table name, source sheet name, and data dimensions.

TableEntry

One registered table and where it came from (CSV path or workbook sheet).

Usage

Point-in-time usage, for display and for the hosted server’s metering.

UsageMeter

Thread-safe in-memory counters for one tenant, checked against a Quota.

Enums

HeaderMode

How the first row of each sheet is interpreted.

QuotaExceeded

Which allowance ran out. The Display text is what the client sees.

Rejection

Why a statement was refused in untrusted mode. The Display text is what the client sees.

SqlTrust

How much the server trusts the client sending SQL.

Constants

ACCEPT_STREAMABLE

Accept value clients should send (re-exported from the shared transport). Accept value a client should send for Streamable HTTP (JSON or SSE responses).

PROTOCOL_VERSION

MCP protocol version reported on initialize.

Traits

TableSource

A supplier of tables. Implementations register whatever they know about into the session and describe it; see the module docs for stateful sources.

Functions

app_router

Build the stateless Streamable HTTP router over the registered tables.

call_tool

Run one tool by name with its JSON arguments and return the result text (unframed).

check_plan

Second decision, on the physical plan: refuse cartesian products. A cross join or a nested-loop join over registered tables is the one query shape a wall-clock limit cannot stop once it runs (DataFusion yields too rarely inside it), so it never starts.

check_untrusted

Parse sql and decide whether an untrusted client may run it against registered tables.

collect_limited

Execute df as a stream, stop at the row/byte caps, and give up at the timeout.

query_runtime

The process-wide tokio runtime queries execute on when a timeout is set: one thread per core, separate from whatever runtime serves clients, so a busy query cannot starve the server. Built on first use.

register_path

Register path by extension. Returns one TableEntry per registered table (workbooks may produce several — one per sheet).

register_sources

Build a registry: register sources in order into ctx and collect their entries.

register_workbook

Open path (xlsx / xls / ods), convert every non-empty sheet to an Arrow RecordBatch, and register each as a DataFusion table. Returns the registered sheets.

run_bounded

Run work under timeout on the query_runtime: the caller’s runtime (HTTP, timers) stays responsive however busy the work is, and on expiry the task is aborted (it stops at its next yield point). Without a timeout, work runs inline.

tools_list

The MCP tools/list result: the five tools with their JSON input schemas.