GPU-accelerated MultiBit Classic wallet password recovery. A custom CUDA kernel does about 11M passwords/sec on an RTX 2060, roughly 20x what the original Python+OpenCL version managed. Reads btcrecover-style token lists and supports save/restore.
Several tools cover different parts of the search space:
| Tool | Language | Rate | Typos |
|---|---|---|---|
multibit_cuda_threads.exe | CUDA C++ | ~11M/s | Yes |
multibit_gpu.py | Python 3 + OpenCL | ~500K-11M/s | No |
multibit_check.py | Python 3 (CPU) | ~67K/s | Yes |
btcrecover.py | Python 2.7 (legacy) | ~54K/s | Yes |
The CUDA tool is the primary workhorse. multibit_check.py is a small, dependency-light
Python 3 checker for portability and for the typo modes it adds; it replaces the vendored
Python 2.7 btcrecover.py, which is kept only for reference.
- CPU producer thread: iterates token list combinations, calls
std::next_permutationto enumerate permutations, assembles password bytes using stack-allocated char arrays (zero heap allocation in the hot loop) - GPU main thread: receives 1M-password batches and runs the MultiBit crypto check (3x MD5, AES-256-CBC, base58 validation), one CUDA thread per password
- The two overlap: GPU checks batch N while CPU fills batch N+1
Python + OpenCL version. CPU generates passwords using Python's itertools.permutations; GPU checks them. Throughput varies with token list complexity (CPU is the bottleneck for complex permutation sets).
Original btcrecover by gurnec, modified with:
- Hot path tweaks:
str.translate()for base58 validation, pre-sliced encrypted blocks - Fast restore: saves
combo_idxalongside the password count so--restorejumps to the right token combination viaitertools.islice(seconds, not hours)
salted = password_bytes + salt (8-byte random salt from wallet file)
key1 = MD5(salted)
key2 = MD5(key1 + salted)
iv = MD5(key2 + salted)
aes_key = key1 + key2 (32 bytes -> AES-256-CBC)
Decrypt the first 32 bytes of the wallet's encrypted section. A valid password produces a Bitcoin WIF private key: first byte in {L, K, 5, Q} and all 32 bytes valid base58 characters.
- NVIDIA GPU, compute capability >= sm_75 (RTX 2060 or newer)
- CUDA Toolkit 13.0
- Visual Studio 2022 Build Tools (MSVC 14.44); note VS 2026 isn't supported by CUDA 13.0
- Windows SDK 10.0.22621.0
See BitCracker/btcrecover-master/multibit_cuda_requirements.txt for full details and known build issues.
pip install -r BitCracker/btcrecover-master/requirements-py3.txt
- Python 2.7
- PyCrypto (recommended for speed):
C:\python27\python -m pip install -r BitCracker/btcrecover-master/requirements-py2.txt
cd BitCracker\btcrecover-master
build_cuda.batThe batch file handles VS 2022 environment setup, CUDA paths, and the TEMP redirect required when the Windows username contains spaces.
.\multibit_cuda_threads.exe --wallet multi.key --tokenlist search46.txt --autosave save46.bin
.\multibit_cuda_threads.exe --restore save46.bin--delimiter <c> changes the character that separates tokens on a line (default
is a space). Use it when a token needs to contain a literal space, e.g. set it to
a comma and write correct horse as a single token.
The CUDA tool also takes btcrecover's typo flags: --typos N plus any of
--typos-capslock, --typos-swap, --typos-repeat, --typos-delete,
--typos-closecase, --typos-insert <charset>. These generate up to N typos per
base password and are meant for a focused pass over a few close candidates, not
the full tokenlist sweep (typo generation allocates and runs well below the
11M/s base rate).
python multibit_gpu.py --wallet multi.key --tokenlist search46.txt --autosave save46.pkl
python multibit_gpu.py --restore save46.pkl --autosave save46.pkl
python multibit_check.py --wallet multi.key --tokenlist search.txt
python multibit_check.py --wallet multi.key --tokenlist search.txt \
--typos 2 --typos-capslock --typos-swap --typos-insert 0123456789
Needs only pycryptodome. Supports --typos-replace and --typos-map, which the
GPU tools don't. Writes a hit to RECOVERED_PASSWORD.txt (owner read/write only).
Kept for reference only; multibit_check.py above supersedes it for MultiBit.
C:\python27\python btcrecover.py --tokenlist search45.txt --wallet multi.key ^
--no-dup --typos 2 --typos-capslock --typos-swap --typos-repeat ^
--typos-delete --typos-closecase --typos-insert %a ^
--utf8 --no-eta --autosave savefileBitCrackerResults --threads 10
Compatible with btcrecover's format:
# Lines prefixed with + are required (must contribute a token)
# Optional lines (no +) may be skipped
+ ^Alpha ^alpha # anchored to first position
+ ^2^one ^2^One # anchored to position 2
Bravo bravo # optional, any position
%0,4d # digit wildcard: 0-4 digits
- Tokens on the same line are mutually exclusive (pick one per combination)
^N^tokenfixes the token at position N (1-indexed)%0,4dexpands to all digit strings of 0-4 digits
All three tools can be interrupted and resumed. The CUDA tool saves combo_idx,
perm_idx, and typo_idx, so restore resumes at the next exact candidate even
inside a large permutation or typo-expanded combination. Old CUDA save files
without perm_idx/typo_idx still load, but only at combo precision.
Save files (.bin, .pkl, savefile*) are excluded from version control.
- The CUDA tool achieves ~11M passwords/sec for token lists with 4-9 free tokens per combination (search46 structure)
- The big win was stack char arrays in the password assembly loop instead of
std::string, which cut out a malloc/free per password and took it from ~3M/s to ~11M/s - For token lists with anchored required tokens and few free tokens (e.g. search47), the CUDA tool completes 76M passwords in ~7 seconds
- The CUDA tool generates typo variants (see
--typosabove); typo generation allocates and runs below the 11M/s base rate, so it's for focused passes over close candidates rather than full sweeps
BitCracker/btcrecover-master/
|-- multibit_cuda_threads.cu # Primary CUDA tool (C++ gen + GPU check)
|-- multibit_cuda.cu # Experimental: GPU-side generation (slower)
|-- multibit_gpu.py # Python + OpenCL GPU tool
|-- multibit_gpu_2.py # Experimental: local memory kernel variant
|-- build_cuda.bat # CUDA build script
|-- multibit_cuda_readme.txt # CUDA tool documentation
|-- multibit_cuda_requirements.txt
|-- multibit_check.py # Standalone Python 3 CPU checker (MultiBit + typos)
|-- tests/ # pytest suite (crypto KATs, typos, tokenlist)
|-- btcrecover.py # Legacy btcrecover entry point (Python 2.7)
`-- btcrecover/
`-- btcrpass.py # Modified: fast restore + base58 tweak
cd BitCracker/btcrecover-master
pip install -r requirements-dev.txt
ruff check multibit_check.py tests/ # lint
python -m pytest tests/ -v # testsThe tests pin the MultiBit key derivation against a known-answer test wallet and
cover the typo generator and tokenlist parser. They need only pycryptodome, so
they run without a GPU (this is what CI runs). See CONTRIBUTING.md.
Licensed under the GNU General Public License v2.0 (see LICENSE).
This project began as a fork of btcrecover
by Christopher Gurnee, which is GPLv2, so this derivative is GPLv2 as well. The
CUDA kernels, the OpenCL GPU tool, the Python 3 checker, and the fast-restore work
are new code (Copyright 2026 Paul Allen); the vendored btcrecover/ tree remains under its
original copyright.