Skip to content

TheDecipherist/compressmcp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

compressmcp

npm GitHub

Lossless JSON compression for Claude Code. Intercepts MCP tool responses and compresses large JSON payloads before they enter Claude's context window, cutting token usage by 40% on average with zero data loss.


Why

Every MCP tool call that returns a database result, API response, or search payload lands verbatim in Claude's context. Verbose JSON field names like transactionId, orderStatus, and repositoryDescription repeat thousands of times across a session, burning tokens on structure rather than content.

compressmcp intercepts those responses and abbreviates the keys using TerseJSON. Claude receives a compact dictionary plus the abbreviated data and reads it just as accurately, but at 40% fewer tokens.


How it works

Three components work together automatically after a one-time install:

Claude asks for data
        │
        ▼
PreToolUse hook ─── detects curl/wget ──► blocks + redirects to mcp__compressmcp__fetch
        │
        ▼ (for direct MCP tool calls)
MCP tool runs, returns JSON
        │
        ▼
PostToolUse hook ─── valid JSON? ──► compress via TerseJSON ──► inject into context
                  └── not JSON?  ──► pass through unchanged

Compressed output format

Every compressed response is three lines:

[compressmcp: 8,366→4,166 tokens (-50%) | lossless]
Keys: {"orderId":"a","orderStatus":"b","customerName":"c","totalAmount":"d",...}
[{"a":"order_001","b":"completed","c":"Customer 1","d":149.99,...},...]

Claude reads the Keys dictionary to decode field names before using the data. The full original structure is recoverable at any time, nothing is dropped, summarised, or truncated.

Compression threshold

Only payloads above 500 estimated tokens (≈2,000 characters) are compressed. Smaller responses pass through unchanged to avoid unnecessary overhead.


Installation

npm install -g compressmcp
compressmcp install

This registers the PostToolUse hook, PreToolUse hook, and MCP server in ~/.claude/settings.json. Restart Claude Code to activate.

compressmcp check     # verify installation status
compressmcp uninstall # remove all hooks and server

Manual build

git clone <repo>
cd compressmcp
npm install
npm run build
compressmcp install

Test results

Token savings benchmark

Tested across five realistic data shapes, each compressed through the real binary pipeline:

Dataset Original Compressed Savings
Orders (100 records) 6,673 tok 4,123 tok -38%
GitHub repos (80 records) 9,719 tok 5,319 tok -45%
Users (200 records) 13,741 tok 7,941 tok -42%
Analytics events (500 records) 28,781 tok 17,656 tok -39%
Products (60 records) 6,748 tok 4,198 tok -38%
Average -40%

Datasets with longer field names compress further (GitHub repos). Datasets with shorter values and more repetition (analytics) hold steady around 39%.

Latency overhead

Hook processing time measured over 10 runs per dataset. Times include Node.js process startup, which dominates the measurement.

Dataset Tokens saved Avg time Est. Claude saving* Net
Orders (100) 2,550 tok 63ms 43ms -20ms
GitHub repos (80) 4,400 tok 61ms 73ms +12ms
Users (200) 5,800 tok 63ms 97ms +34ms
Analytics (500) 11,125 tok 69ms 185ms +116ms
Products (60) 2,010 tok 62ms 34ms -29ms

* Based on Claude Haiku input throughput (~60k tokens/sec). Positive Net = compression saves more time than it costs.

The ~60ms baseline is process startup — a one-time cost per tool call. The real payoff isn't speed: it's context space. Every token compressed is a token freed from Claude's context window, leaving more room for code, conversation, and reasoning on every subsequent turn in the session.


Live status bar

compressmcp includes a Claude Code status bar that shows real-time compression stats for your active session.

compressmcp status bar

The bar displays left to right:

Section Example Description
Context bar 22% | 43K/200K How full Claude's context window is (colour: green → yellow → red)
Model sonnet Active model (opus / sonnet / haiku)
Token savings -4.7M tok · 40% Tokens removed this session + compression efficiency percentage
Call count 907 calls Number of compressed responses so far
Plan usage 5h [9%] 7d [17%] Claude plan rate-limit utilisation (5-hour and 7-day windows)
Branch main Current git branch

Enable the status bar

Add the statusLine command to ~/.claude/settings.json (done automatically by compressmcp install):

{
  "statusLine": "compressmcp --status"
}

The command reads live context and rate-limit data piped in by Claude Code, combines it with compression stats from the current session's JSONL log, and writes the formatted bar to stdout.


Safety

Compression is strictly lossless. The test suite includes:

  • Lossless roundtrip tests, every compressed payload decompresses back to byte-identical JSON
  • No-passthrough-corruption tests, non-JSON content passes through untouched
  • End-to-end integration tests, spawns the real compiled binary, pipes actual HookInput JSON, asserts on stdout
  • Claude comprehension test, verifies the compressed format (header + dictionary + data) is correctly structured for Claude to decode
npm test

262 tests across unit, safety, integration, and benchmark suites.


License

MIT