@string-utility-belt/mcp

@string-utility-belt/mcp

An MCP server that exposes String Utility Belt's 247 string-transformation utilities — and its pipeline engine — as tools an AI agent can call directly, with no browser involved.

Install

Nothing to clone or build: clients run it from npm with npx (Node.js 20 or later).

Claude Code

claude mcp add subelt -- npx -y @string-utility-belt/mcp

Or install it as a plugin, which pins the server to the release the plugin ships with (plugin README):

/plugin install string-utility-belt --marketplace String-Utility-Belt/string-utility-belt

Claude Desktop, Cursor and other MCP clients: add this to the client's MCP configuration (for Claude Desktop, claude_desktop_config.json), then restart the client:

{
  "mcpServers": {
    "subelt": {
      "command": "npx",
      "args": ["-y", "@string-utility-belt/mcp"]
    }
  }
}

It is also on Smithery and in the official MCP Registry as com.stringutilitybelt/mcp. Then ask your agent for something like "base64-decode this and pretty-print the JSON": it finds the utilities with list_utilities and runs them with run_utility or run_pipeline.

Tools

• list_utilities — { category?, query?, limit? } → { items: [{ id, name, category, description }], total, categories }, best matches first. query is a set of words that must all match (in any order) somewhere in the id, name, tags, aliases or description. Start here.
• describe_utility — { id } → a utility's full doc: params (kind, default, options/bounds, description), what it accepts/produces, tags, aliases, worked examples, and the runtime capabilities it needs.
• run_utility — { id, input, inputEncoding?, params? } → { output, outputEncoding? }. Text output is returned as-is; JSON output is pretty-printed text; byte output is base64 with outputEncoding: "base64". Params are validated against the utility's spec. A failing utility comes back as isError: true with the message, never a thrown exception.
• run_pipeline — { steps?, share?, input, inputEncoding? } → { output, outputEncoding?, errors, timings, skipped, halted } for a whole pipeline (the v3 schema: utility steps, branches, macros, "run on each" steps, conditions, per-step error policy). Pass exactly one of steps ([] is the identity pipeline) or share (a #/p/<payload> link, a URL containing one, or the bare payload). Unknown utility ids and invalid params on enabled steps are rejected up front.
• detect_format — { input, inputEncoding? } → { candidates: [{ format, confidence, note }] }, best first (JSON, YAML, base64, JWT, gzip, …).

inputEncoding is "text" (default), "base64" (standard or URL-safe, padding optional), "hex" (both decode to raw bytes first) or "json" (parse first); detect_format takes "text", "base64" or "hex".

Resources

• subelt://utilities — the full catalog as JSON (id, name, category, description, tags, aliases for every utility).
• subelt://utility/{id} — one utility's full documentation (same shape as describe_utility), via a resource template with id completion.

Limits and isolation

• Input is capped at 1 MB per call (a share link at 1M characters).
• Output past 1M characters is cut off and flagged truncated: true with the untruncated fullLength.
• A pipeline is capped at 100 steps (nested steps counted).
• Every call is stopped after 20 s — set SUBELT_MCP_TIMEOUT_MS to change that.
• Utility code runs in pooled child processes (up to 4, 512 MB heap each), not in the server: a call that blocks (a catastrophic regex, bcrypt at cost 31) is killed at the deadline, one that exhausts memory fails with "ran out of memory", and either way the server keeps answering. Runners exit with the server, and kill themselves if it dies abruptly.
• Utilities that run arbitrary user JavaScript (custom_js) are always refused — LLM-driven code execution is out of scope for this server. The two utilities that need a browser DOMParser (xml_to_json, html_table_to_csv) fail with an explanatory error.

Privacy Policy

The server runs entirely on your computer. Our full privacy policy is at stringutilitybelt.com/privacy/; for this server it comes down to:

• Data collection. It receives only the arguments your MCP client sends in each tool call (the input text, utility ids, parameters, pipeline steps) and returns results only to that client. It contains no analytics, telemetry or ads.
• Use and storage. Inputs are processed in memory, by the server and its child processes, to produce the result you asked for. Nothing is written to disk. Diagnostics, such as a failed start or a runner that ran out of memory, go to standard error, which your MCP client may keep in its own logs.
• Third-party sharing. None. The server makes no network requests, so nothing you give it reaches us or anyone else. (Your MCP client may download the icon listed in the server's metadata from stringutilitybelt.com to display it; the server itself does not.)
• Retention. Nothing is kept once a call returns.
• Contact. [e-mail].

Building

From the repo root (no separate install needed — it reuses the root node_modules):

npm run build:mcp

This produces a single standalone file at packages/mcp/dist/server.mjs (Node 20+, every dependency bundled in — only Node's own built-ins are external). It is an SSR build, so dependencies resolve to their Node variants; the same file doubles as the job-runner entry (--job-runner).

Running a local build

To try changes from a checkout, point the client at the built file instead of the npm package:

claude mcp add subelt -- node /absolute/path/to/string-utility-belt/packages/mcp/dist/server.mjs

Development

# Unit + integration tests (in-process via InMemoryTransport, plus the process pool)
npx vitest run packages/mcp/src --maxWorkers=2

# Build, then also run the stdio smoke tests against the real binary
npm run build:mcp
npx vitest run packages/mcp/src/bin.smoke.test.ts --maxWorkers=1

Source and issues

@string-utility-belt/mcp is developed in the String Utility Belt repository on GitHub, alongside the web app and the other integrations. Report bugs or request features in its issue tracker. MIT licensed.