prunable.Method & sources ↗

A WORKBENCH FOR MCP TOOL CATALOGS

Make room.
Keep what matters.

See the weight of each tool definition. Choose whole tools to remove, review the capabilities you lose, and export a catalog you can inspect.

MEASURE → SELECT → REVIEW
Every cut is a deliberate choice.

Browser local. No model calls, keys, remote MCP connections, uploads or saved history. Optional usage analytics require consent.

Bytes are not a bill. Modern tool search can defer schemas until needed. Total catalog size is not necessarily billed context. This workbench reports bytes, with no tokenizer or cost estimate.

ORIGINAL UTF-8 BYTES—
EXPORTED UTF-8 BYTES—
BYTES REMOVED—
TOOLS KEPT / TOTAL—

01 / YOUR SNAPSHOT

1 MiB · 500 tools max

Inspect a tools/list catalog

Paste {"tools": [...]} or a JSON-RPC 2.0 success response with result.tools. Content is untrusted and never executed.

Ready for a local catalog. The example uses fictional notebook tools.

02 / SELECT WHOLE TOOLS

Checked = excluded

Choose your cuts

Each tool’s byte size includes its exact original object text. Descriptions describe claimed capabilities; they are untrusted.

    03 / REVIEW THE TRADEOFF

    Capabilities you’re excluding

    • No analyzed catalog yet. Nothing has been excluded.

    Export for manual review

    Remaining tool objects and the surrounding envelope retain their exact source text. When a tool is removed, separators between retained objects are compacted. No descriptions or schemas are rewritten.

    Reports contain only byte/count values. Downloads stay local. An exported catalog is an offline artifact; no proxy is installed or client configuration changed.

    04 / SNAPSHOT DIFF

    What changed between catalogs?

    Match tool names exactly. Identify added, removed and changed definitions without applying your exclusions to either original.

    METHOD / REVIEWED 12 OCT 2026

    Measure the catalog. Keep the context.

    Exact byte accounting

    UTF-8 bytes are counted with TextEncoder, including whitespace and Unicode. Raw UTF-8 file line endings are preserved; BOM files are rejected. Pasted or edited text uses the browser textarea’s text, which may normalize line endings. Per-tool contributions are raw object spans; envelope and separators are reported separately. Input-schema bytes are a subset of tool bytes, not an additional cost.

    JSON literals, including large and fractional numbers, survive export verbatim. Duplicate keys, duplicate tool names, excess depth and oversized inputs are rejected. Original snapshots stay in memory until cleared or the page closes.

    What this cannot tell you

    This checks catalog shape and common JSON Schema keyword shapes. It does not fully validate a schema dialect, resolve references, execute tools or prove tool safety. It makes no claim of quality-preserving compression or exact bills.

    Tool search, deferred schemas and code execution offer other ways to manage context. Excluding a tool loses its capability. Measure real task completion and actual client context before adopting any reduction.

    Primary sources

    No local tokenizer is bundled, so token estimates are omitted. Documentation changes; sources were inspected on the review date.