FM-39 · Seed phrase concealmentDistribution: unlimited. Copy freely.
Bitcoin Seed BookSelf-custody field manual

Split a seed across places

Home » Resources » Split a seed across places

One backup in one place has two ways to fail. A fire takes it and you have nothing. A burglar takes it and they have everything. Splitting the backup across places is how you trade those two risks against each other. Some ways of splitting do that well. The obvious one, cutting the phrase into pieces with scissors, does it badly.

Don't cut the phrase

Take the public example seed:

pizza educate fat kiss domain venue … (this phrase is public: never use it for real funds)

Cut a 12-word phrase into three pieces of four words and you have made the backup more fragile, because all three pieces are now needed. To let any two pieces restore it, each piece has to carry eight words (words 1 to 8, 5 to 12, and 1 to 4 plus 9 to 12). Now a thief who finds one piece is missing four words in known positions. That is 2,048 to the fourth power, about 1.8 × 1013 candidates, of which one in 16 passes the checksum. Our brute-force page calls four missing words "a serious investment, still not safe".

A 24-word phrase cut the same way leaves eight words missing: about 280 candidates after the checksum. That is out of reach today. It still hands 16 of the 24 words, 176 bits, to whoever opens one envelope, and each piece is plainly a list of seed words. The schemes below give a thief holding one piece nothing at all.

Threshold schemes: Shamir shares

Shamir's secret sharing splits a secret into N shares so that any T of them rebuild it and fewer than T reveal nothing. The usual choice is 2-of-3. Two standards apply it to wallet seeds.

SLIP-39

SLIP-39 is SatoshiLabs' standard. Each share is a list of words from its own 1,024-word list: 20 words for a 128-bit secret, 33 for 256 bits. The last three words of every share are a checksum that catches any error in up to three words. The spec says that "knowledge of fewer than the required number of parts does not leak information about the master secret."

It also allows two levels: groups, each with its own threshold, and a threshold of groups. Up to 16 groups of up to 16 shares. You could require two of three groups, where one group is you (any two of your three shares) and another is a lawyer holding one.

Three things to know before you choose it:

  1. It is not BIP-39. The spec says the two are "for the most part" incompatible. A BIP-39 phrase can be converted, but every share then grows to 59 words, and shares can't be turned back into a BIP-39 phrase. Plan to make a new wallet with SLIP-39 from the start and move the coins.
  2. It has its own passphrase option, and, like BIP-39, there is no way to check that the passphrase you typed was the right one.
  3. Trezor implements it as "Multi-share Backup", up to 16 shares in any T-of-N. Trezor's own FAQ says the configuration can't be changed once it is set up; to add a share you make a new backup.

Codex32 (BIP-93)

Codex32 does the same job with paper and a pencil. BIP-93 states the aim: "to compute and verify checksums, and to split and recover seeds, entirely using pen and paper." The secretcodex32.com booklet supplies the lookup tables, worksheets and paper volvelles (rotating wheels) to do it.

  • A 128-bit seed fits in 48 characters, written in the bech32 alphabet.
  • The threshold is 2 to 9; a seed can be split into up to 31 shares.
  • "Any number of shares less than the threshold leaks no information about the secret."
  • The checksum can correct up to four wrong characters or eight unreadable ones.

The cost is time. The booklet warns that a mistake on the checksum worksheet may go unnoticed until you have spent "upwards of an hour" on it. Codex32 encodes a BIP-32 master seed directly and is not a BIP-39 phrase, so as with SLIP-39 you start a new wallet. BIP-93's authors "do not recommend interconversion".

XOR splitting: every part needed

XOR splitting is N-of-N. You make N minus 1 random phrases of the same length as the seed. The last part is the seed's entropy XOR-ed with all of them, so the XOR of all N parts gives back the seed. Each part is written down as an ordinary BIP-39 phrase with its own checksum.

The maths is the one-time pad. With all N parts you have the seed. With N minus 1 parts you have random numbers that say nothing about it, because every possible seed fits them equally well. That is the strength. The weakness is the same fact: lose one part and the seed is gone, for you as well as the thief. XOR protects against theft and does nothing for loss. If you use it, each part needs its own backup, or a copy in a second place, which brings back the problem you were splitting to avoid.

Do the XOR by hand or on an offline device, never on a web page or an online tool. Then restore the seed from the parts on a second offline device and check that it gives the same wallet before you fund it.

The simplest split: words here, passphrase there

Before any of this, consider two places and no new standard. Keep the BIP-39 words in one place and the passphrase in another. Either alone opens an empty wallet. Like XOR, this is two-of-two: lose the passphrase and the words are useless. So each half still needs a copy. For many people this is enough, and it's the split this site's method is built around.

Multisig is the other route. It splits keys, not a seed: a 2-of-3 multisig wallet has three separate seeds, and spending needs signatures from two. It also needs every backup to record the wallet's configuration, which is a subject for its own page.

Hiding is not encrypting. The ciphers are public and a seed phrase carries its own checksum, so a list someone recognises as a seed can be tried against every cipher in minutes. Add a BIP-39 passphrase to the wallet and keep it somewhere else.

How many places is enough

A T-of-N scheme survives N minus T losses and T minus 1 thefts. For 2-of-3:

  • Lose one share in a flood: two remain, you restore.
  • A thief takes one share: they hold one, which is nothing, and you still have two.
  • Both at once: you are down to one share and can't restore. A thief who only photographs a share leaves it with you, and then 2-of-3 survives the loss as well.

A 3-of-5 survives two losses or two thefts, and one of each. That is the case for going bigger. The case against is everything else:

Scheme Survives Places to keep
Words + passphrase no loss without copies 2
Shamir 2-of-3 1 loss or 1 theft 3
Shamir 3-of-5 2 losses or 2 thefts 5
XOR 3-of-3 no loss; a thief needs all 3 3

Every place is a person you trust or a trip you make. Each share should be tested when it's made and checked now and then, and five places means five trips. Each one needs to hold up to fire and water (metal backups) and to show if it has been opened (tamper-evident backups). Places that share a building, a town or a flood plain count as one place. Write down, apart from the shares, what the scheme is and where the pieces are, so whoever recovers the wallet after you knows that three of five are needed.

What to do

  1. Don't cut a phrase into pieces.
  2. If two places are enough, keep the words in one and the passphrase in the other, with a copy of each.
  3. If you want a threshold, use SLIP-39 or Codex32, 2-of-3 unless you have a reason for more. Start a fresh wallet with it.
  4. Use XOR only when every part also has its own backup.
  5. Put the shares where one event can't reach two of them.
  6. Restore from the minimum number of shares before you fund the wallet, then on a schedule.

We will never ask for a seed word or a passphrase. Never type one into this site or any tool it links to.

Sources