SPL Token instructions use a one-byte tag followed by fixed fields, not an Anchor discriminator. Transfer is tag 3 with a u64 amount; TransferChecked is tag 12 with an amount and a decimals byte, and it is the only transfer that lists the mint among its accounts. Account order, not names, tells you which account is which.
Tags, accounts and data#
The program is TokenkegQfeZyiNwAJbNbGKPFXCWuBvf9Ss623VQ5DA. These are the instructions you meet most often, with the single-owner account order:
| Tag | Instruction | Accounts, in order | Data after the tag |
|---|---|---|---|
| 3 | Transfer | source, destination, owner | u64 amount |
| 4 | Approve | source, delegate, owner | u64 amount |
| 7 | MintTo | mint, destination, mint authority | u64 amount |
| 8 | Burn | account, mint, owner | u64 amount |
| 9 | CloseAccount | account, destination for the lamports, owner | none |
| 12 | TransferChecked | source, mint, destination, owner | u64 amount, u8 decimals |
| 15 | BurnChecked | account, mint, owner | u64 amount, u8 decimals |
| 17 | SyncNative | the wrapped SOL account | none |
| 18 | InitializeAccount3 | account, mint | owner Pubkey |
When the owner is a multisig, the owner's slot holds the multisig account and the signers follow after the listed accounts. All amounts are in the mint's base units; the decimals byte in the checked variants tells you where the decimal point goes.
Transfer or TransferChecked: where the mint is#
A plain Transfer names a source token account and a destination token account. It does not name the mint, so from the instruction alone you cannot tell which token moved. You need the token account's data, which holds its mint, or a mapping you built earlier from the instruction that created that account.
TransferChecked adds the mint as its second account and the decimals as a final byte, and the program refuses the transfer if either does not match. For a decoder it is far easier to read: token, amount and scale are all inside the instruction. Wallets and most newer programs send TransferChecked for that reason, and the Token-2022 program requires the checked form for mints with certain extensions, which Token-2022 extensions explains.
Three newer tags since p-token#
On mainnet the token program at the same address now runs p-token, a reimplementation described in SIMD-0266. Its feature gate activated at slot 419,472,000, in epoch 971 (13 May 2026). It keeps every existing instruction and layout and adds three:
- tag 38, WithdrawExcessLamports, with the same value Token-2022 uses;
- tag 45, UnwrapLamports, which moves lamports out of a wrapped SOL account directly;
- tag 255, Batch, which carries several token instructions in one call.
Batch changes how you parse. Its data is a sequence of sub-instructions, each with an account count, a data length and its own tag, and its accounts are the concatenation of theirs. A decoder that knows only tags 0 to 24 will either skip it or misread it. Programs mostly use Batch through CPI, so you will meet it more in execution metadata than at the top level.
What two real blocks contained#
Two consecutive mainnet blocks on 18 November 2026, counted by tag:
| Instruction | Top level, in the raw transactions | Inside other programs, as CPIs |
|---|---|---|
| TransferChecked (12) | 19 | 471 |
| Transfer (3) | 0 | 16 |
| CloseAccount (9) | 235 | 7 |
| SyncNative (17) | 133 | 3 |
| InitializeAccount3 (18) | 0 | 250 |
| UnwrapLamports (45) | 0 | 1 |
At the top level, token instructions were mostly housekeeping: closing accounts and syncing wrapped SOL around swaps. The transfers themselves happened inside DEX and router programs.
Why most token transfers never reach a shred decoder#
The table above is the practical point. A swap's token movements are CPIs made by the swap program, and inner instructions exist only after execution. A shred decoder sees the token instructions the user signed directly, such as a wallet-to-wallet TransferChecked or the CloseAccount at the end of a swap, and reads the swap's intended amounts from the swap instruction instead. If you need every token movement, you need executed data as well.
Questions about token instructions#
Why does my plain Transfer have no mint?#
Because the instruction never carried one. Look up the source token account's mint, from account data or from the instruction that created it.
Are SPL Token amounts Borsh-encoded?#
The program packs its own layout: a tag byte, then fixed little-endian fields. For a u64 amount the bytes match Borsh, but optional values do not, so do not decode token instructions with a generic Borsh reader. The Borsh layout rules explain the difference.
Is Token-2022 the same layout?#
Its base instructions use the same tags and layouts, at a different program ID, plus extension instructions of its own.
What token instructions cannot tell you#
A top-level token instruction is the request, not the result. It cannot tell you whether the transfer succeeded, what the balances became, or which owner a token account belongs to if you have not seen it created. And most token activity on Solana is not top-level at all.
Decoding them in practice#
The SPL Token transfer recipe decodes Transfer and TransferChecked in code and checks them against balance changes, the Token-2022 recipe covers the newer program, and associated token accounts explains the addresses these instructions point at. Instruction indexes themselves are covered in compiled instructions; to run the decoder on a feed, try a first decode.
Sources#
- solana-program/token interface/src/instruction.rs, commit 3d53b93: tags, account order and data (checked 2026-11-18)
- SIMD-0266 p-token: Efficient Token program (batch 255, unwrap_lamports 45, withdraw_excess_lamports 38) (checked 2026-11-18)
- Agave v4.4.1 feature-set: replace_spl_token_with_p_token; mainnet feature account activated at slot 419,472,000 (checked 2026-11-18)
- Mainnet getBlock slots 470,901,815 and 470,901,816 (top-level instructions and meta.innerInstructions) (checked 2026-11-18)
Raw shreds on your own server
Ultra Low Latency Shreds is $200 and Standard Shreds is $100. You pay in crypto, tell us which IPv4 address and UDP port should receive the stream, and we wire it up by hand.