MCP server de Sezzlee XML

@sezzlee/xml-mcp

A read-only, sandboxed MCP server that reads local XML documents. It builds on @sezzlee/file-core; it does not depend on @sezzlee/core and imports nothing from packages/lab/xml-lab.

Documentation: https://docs.sezzlee.app/docs/xml-mcp/introduction — a tutorial, task guides, and the tool, error-code and limit reference generated from this package.

Quick start

The root is the only argument, as an absolute path; there is no environment variable for it.

{
  "mcpServers": {
    "xml": {
      "command": "npx",
      "args": ["-y", "@sezzlee/xml-mcp", "/path/to/xml/root"]
    }
  }
}

With Claude Code:

claude mcp add xml -- npx -y @sezzlee/xml-mcp /path/to/xml/root

Readable extensions: .xml, .xsd, .xhtml, .svg, .csproj, .props, .targets, .config, .resx. An extension like .config is not a guarantee of XML — a parse error there is expected behavior. Requires Node.js 22 or 24 on macOS (x64, arm64), glibc Linux (x64, arm64) or Windows x64.

Tools

Tool	What it does
list_documents	Lists readable XML documents under the root without parsing them
describe_document	Document element, every namespace with a stable alias, structure counts, repeated-element candidates, and an address read_node takes
read_node	A bounded, ordered slice of the tree as flat depth-first records
find_in_document	Literal text in text nodes, attribute values or both
select_xpath	One XPath 1.0 expression, exactly as written, with a typed result
project_records	A repeated element as rows and named columns, with a status per cell
aggregate_document	Counts, distinct counts and, with numericMode: binary64, sums and averages per group

Every argument, error code and limit is listed in the reference.

A document over the 8 MiB resident ceiling is read in a reduced-capability chunked mode: select_xpath, find_in_document and aggregate_document are unavailable there, while project_records still works.

Configuration

createXmlMcpServer(root, options) takes two optional settings; the CLI always uses the defaults.

Option	Default	Meaning
documentCacheSize	4	Documents kept resident in memory at once (1–64)
maxConcurrentListings	4	Concurrent list_documents calls

createXmlMcpServer(root, {
  documentCacheSize: 4,
  maxConcurrentListings: 4,
});

documentCacheSize (S) is the single knob; the worker pool's capacity is derived as W = 2S, and the invariant W >= 2S-1 is pinned by a test — the worker keeps a document alive while the store still holds it, so its map must fit every store entry plus the one being adopted. This is a budget, not enforcement: worker.resourceLimits bounds the JS heap, not WASM linear memory, so exceeding the budget is a process crash rather than a clean resource_limit error. Measured cost is 9.25-10.08x the source byte count, so roughly 81 MiB resident per document at the 8 MiB ceiling; the default S=4 is safe on the smallest supported host.

Rules

Parsing happens in the worker. The main process only ever holds a serializable handle; a WASM pointer never crosses the boundary. No disposal hook was needed on @sezzlee/file-core's document store because of this.

Layout follows the worker/host boundary. engine/ is the worker-side graph and never names host/ or tools/; model/ and primitives/ are the shared vocabulary between the two sides. index.ts, cli.ts, xml-worker.ts and server.ts stay at the root — the tarball entry point, bin, the CI tarball check and the ../package.json read respectively tie them there.

The worker entry never imports the host surface. src/xml-worker.ts only sees node:worker_threads, node:buffer, libxml2-wasm and a type-only protocol; a lint rule enforces this. A worker returns a code string, and the host side constructs the error object.

DOCTYPE is refused before parsing. The prolog scanner runs in the main process; doc.dtd is only the second check. Measured: XML_PARSE_NO_XXE does not block the internal DTD subset.

The worker's stdout never mixes into the parent's fd 1. It is isolated with stdout: true; on stdio MCP, a single stray line breaks JSON-RPC.

diag is off in production and cannot be turned on through an environment variable: it costs 24.9% and its raw report carries an engine pointer.

Unsupported encodings are refused before parsing. The prolog scanner applies the XML 1.0 Appendix F four-byte autodetection; the UCS-4 and EBCDIC families get unsupported_encoding. Measured: the previous scanner missed DOCTYPE in these families.

libxml2-wasm is pinned to exact 0.7.2. A caret would silently invalidate the F0-01 integrity record; the CI tarball checker enforces this.

Development

pnpm turbo run build --filter=@sezzlee/xml-mcp
pnpm turbo run test --filter=@sezzlee/xml-mcp
pnpm turbo run check-types --filter=@sezzlee/xml-mcp

Run tests through Turbo, not pnpm --filter @sezzlee/xml-mcp test: the bare filter skips dependsOn: ["^build"].

A change to a tool's arguments, an error code or a limit changes the generated reference in apps/docs; run pnpm --filter @sezzlee/docs gen and commit the result, or pnpm turbo run validate --filter=@sezzlee/docs fails.

Large-document tests are opt-in and skipped by default; set SEZZLEE_XML_LARGE=1 to run them.