msgpack/msgpack-node: MessagePack implementation for Node.js

node-msgpack is an addon for Node.js that serializes and de-serializes JavaScript values with MessagePack. Packed output is a Buffer and is typically much smaller than JSON.

Version 3.3 requires Node.js 18+, vendors msgpack-c c-7.0.2, unpacks 64-bit integers outside Number.MAX_SAFE_INTEGER as bigint, accepts optional pack type/family hints, can unpack maps and arrays lazily (unpack(buf, { lazy: true })), and applies write backpressure on Stream.send. See SECURITY.md.

Usage

const assert = require('assert');
const msgpack = require('msgpack');
const o = { a: 1, b: 2, c: [1, 2, 3] };
const b = msgpack.pack(o);
const oo = msgpack.unpack(b);
assert.deepEqual(oo, o);

pack() accepts any JSON-like value plus Node Buffers, Dates, and bigint values in the int64/uint64 range. unpack() consumes a Buffer and returns a JavaScript value, or null if the buffer is a truncated (incomplete) MessagePack object. Oversized array/map/string bombs throw.

A streaming helper wraps a readable socket and emits msg, plus error when a packet cannot be unpacked (the offending buffer is dropped). send() packs and writes; it returns the boolean from the underlying write(), or false if the message was queued because a previous write returned false and drain has not fired yet. drain is re-emitted from the underlying writable onto the Stream. At most 1024 messages may wait in that queue; a further send() throws. Extra write arguments (encoding, callback) are forwarded on an immediate write. On underlying error / close / end, queued messages are dropped and Stream emits error if any were unsent:

const msgpack = require('msgpack');
const ms = new msgpack.Stream(socket);
ms.on('msg', (m) => {
  console.log('received', m);
});
ms.on('error', (e) => {
  console.error('bad packet', e.message);
});
ms.on('drain', () => {
  / underlying writable is ready for more send() calls /
});
ms.send({ hello: 'world' });

Type mapping (3.0)

Packing:

• undefined / null → nil
• boolean → bool
• finite integers (number or bigint in the 64-bit range) → uint/int
• bigint outside uint64/int64 → throws
• other numbers → float64
• string → str (UTF-8)
• Date → str (ISO 8601, toISOString()), at any nesting level
• Buffer → bin
• Array → array
• objects with a toJSON() method → whatever toJSON() returns, at any

• other objects → map of every own enumerable key; numeric keys are packed as

• functions, circular refs, and nesting deeper than 512 throw

Unpacking:

• nil → null
• bool / float → JS boolean / number
• int whose magnitude ≤ Number.MAX_SAFE_INTEGER → number (a uint64 of

• int whose magnitude > Number.MAX_SAFE_INTEGER → bigint
• str → string
• bin → Buffer
• array / map → Array / Object
• ext → throws

unpack.bytes_remaining is the number of unused trailing bytes after the last successful (or attempted) unpack. Stream uses that to splice leftover data.

unpack(buf, { lazy: true }) wraps maps as objects with accessor own-properties and arrays as array-likes with indexed accessors. Nested values are not converted until they are read, which is useful for large payloads when only a few keys are needed. The decoder copies buf so later reads do not depend on the caller's backing store (transfer / detach is safe). JSON.stringify and util.inspect materialize via toJSON / inspect.custom. Lazy arrays are not real Arrays (Array.isArray is false); pack() still round-trips them because it calls toJSON. Primitives unpack eagerly even when lazy is set. __proto__ and constructor keys stay own properties, same as eager unpack.

Pack type hints (3.1)

pack(value, options) takes an optional last-argument options object when there are exactly two arguments and that object own-enumerates only type, family, and/or interpret. Extra keys, a one-argument { type: ... } value, and pack(1, 2) still pack as values / an array.

msgpack.pack(123, { type: 'fixint' });          // 0x7b
msgpack.pack(123, { type: 'uint8' });           // 0xcc 0x7b
msgpack.pack(Math.PI, { type: 'float32' });     // 0xca + 4 bytes
msgpack.pack(buf, { family: 'bin' });
msgpack.pack(1.5, { family: 'int' });           // throws
msgpack.pack(500, { type: 'uint8' });           // throws
msgpack.pack(
  [
    { data: Math.PI, type: 'float64' },
    { data: 3.14, type: 'float32' },
  ],
  {
    interpret(item) {
      return { data: item.data, type: item.type };
    },
  },
);

type forces that MessagePack type (fixint, uint8…uint64, int8… int64, float32/float64, fixstr/str8/str16/str32, bin8/bin16/ bin32, nil/true/false). family (int, float, str, bin) picks a compact encoding in that family. If both are set, type wins. Out-of-range values throw cannot pack value as .

interpret is used when packing an Array. Each element is passed to interpret(item), which must return { data } and may also set type / family for that element. Nested interpret on the returned object is ignored.

Default packing is unchanged when no recognized options object is passed.

Limits

• array/map length ≤ 1,000,000
• str/bin/ext length ≤ 32 MiB
• nesting depth ≤ 512 on both pack and unpack

Building, installation, testing

npm install
npm test
npm run coverage

Needs a C/C++ toolchain and Python (node-gyp). GitHub Actions runs Node 18/20/22 on Ubuntu and macOS. npm run coverage instruments JavaScript with c8 and the native addon with gcov, and fails under 95%. Gates and remaining uncovered lines are documented in COVERAGE.md.

Command Line Utilities

Two utilities convert between JSON and MessagePack on stdin/stdout: bin/json2msgpack reads JSON and writes MessagePack, and bin/msgpack2json reads MessagePack and writes JSON. Both are installed on PATH when the package is installed globally.

echo '[1, 2, 3]' | ./bin/json2msgpack | xxd

00000000: 9301 0203                                ....

Piping the two together round-trips a value:

echo '[1, 2, 3]' | ./bin/json2msgpack | ./bin/msgpack2json

[1,2,3]

echo '{"hello":"world"}' | bin/json2msgpack | bin/msgpack2json

{"hello":"world"}

msgpack2json prints one JSON value per line and consumes every complete message in its input. Both exit non-zero on invalid or truncated input.

Benchmarks

npm run bench

or equivalently:

node test/benchmark/benchmark.js

The benchmark serializes the object {'abcdef': 1, 'qqq': 13, '19': [1, 2, 3, 4]} 500,000 times through four paths, after a warm-up pass. A representative run:

node       v20.20.2
v8         11.3.244.8-node.38
platform   linux 6.12.76-linuxkit (arm64)
cpu        arm64 (model not reported) x 8, 7.8 GiB RAM
data       {"19":[1,2,3,4],"abcdef":1,"qqq":13}
iterations 500,000
JSON.stringify()                     179 ms  (0.18 s)
JSON.parse(JSON.stringify())         346 ms  (0.35 s)
msgpack.pack()                      1021 ms  (1.02 s)
msgpack.unpack(msgpack.pack())      1585 ms  (1.59 s)

Measured on 2026-09-10: Node.js v20.20.2, Debian 12 (bookworm), Linux 6.12.76-linuxkit aarch64 container (8 vCPUs, 7.8 GiB RAM). Numbers are for that machine and object shape only — re-run npm run bench on your own hardware before drawing conclusions.

On small objects like this one, V8's native JSON codec is faster than crossing the JS/C++ boundary per call; msgpack's advantage is payload size (20 bytes here versus 36 for the JSON text) and its ability to carry binary data without base64. Large Buffer payloads and batched (single-call) packing shift the comparison considerably.

License

This addon is BSD-3-Clause (Copyright (c) 2010, Peter Griess); see LICENSE.

The vendored MessagePack C library in deps/msgpack/ is **Boost Software License 1.0**, as shipped by msgpack-c c-7.0.2; see deps/msgpack/LICENSE. msgpack-c was Apache-2.0 through 1.2.x and relicensed to BSL-1.0 in release 1.3.0 (2015-11-21), so the Boost text is the correct license for these files:

• relicensing discussion:
• msgpack-c CHANGELOG.md, 1.3.0: "Change license from Apache 2.0 to Boost