Codync
무료 오픈소스 1:1 대안, Grok Bot과 Muse 대신
저장소 정보는 아직 수집하지 않았습니다.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
Send big files straight to a friend's machine, encrypted end to end. Also a channel between two Claude Code sessions.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다.
README 전문 보기 ↗Send files and messages between two machines, yours and a friend's. No accounts, no port forwarding, end-to-end encrypted.
ferry send ./build.tar.zst
It prints a code like 7-acorn-tulip-otter. On the other machine:
ferry get 7-acorn-tulip-otter
That is the whole thing. The file goes straight from one machine to the other as fast as the slower connection allows, survives a dropped Wi-Fi, resumes if interrupted, and is checked byte for byte on arrival.
ferry pair # prints a code; the friend runs: ferry pair <code>
ferry send sam ./photos # from now on, no codes
ferry msg sam "pushed branch x, have a look"
ferry inbox
ferry wait --from sam --timeout 10m
Messages reach a friend who is offline too: they wait, sealed, in a
mailbox until the friend is next online. Files need both machines online;
ferry up keeps a machine reachable in the background.
ferry is an MCP server:
claude mcp add ferry -s user -- ferry mcp
Two Claude Code sessions on two machines can then message each other,
wait for replies, and hand over files and patches. The tools are
send_message, read_messages, wait_for_message, send_file,
list_files_received, fetch_file, list_contacts and channel_status.
Shell-only agents get the same through the CLI: every command takes
--json, writes one JSON event per line to stdout and progress to stderr,
never prompts, and exits with a meaningful code (ferry help exit-codes).
ferry help --llm prints a one-page manual written for a model.
What arrives through ferry is data from someone else. ferry never runs it, fetched files lose their execute bit, and the MCP server tells the model to treat received text as untrusted.
One static binary for Linux and macOS:
go install github.com/ZahakJ/ferry/cmd/ferry@latest
To reach each other across the internet, two machines need a rendezvous server to introduce them, and a way through their routers. Run your own rendezvous (deploy/fly has the steps for fly.io; any machine with a public address works) and point both machines at it:
ferry config rendezvous /dns/rv.example.org/tcp/4001/p2p/12D3KooW...
If a router blocks the direct connection, forward one UDP port to the machine behind it and tell ferry to use it:
ferry config port 40000
ferry check
ferry check says what works and what does not: the rendezvous, your
public address, the port.
STUN (how each machine learns its public address) defaults to a public
server. To use your own, set FERRY_STUN=stun:stun.example.org:3478, or
"stun": ["stun:…"] in config.json in ferry config's folder.
ferry pipe connects stdin/stdout to a TCP port on a contact's localhost,
over the same direct, end-to-end encrypted link, with no open ports on
either side. On the machine to reach, allow exactly the ports you mean
(nothing is allowed by default; every pipe is logged):
ferry up --allow-pipe 22
On the other, after pairing:
ssh -o ProxyCommand='ferry pipe box 22' me@box
# or in ~/.ssh/config: Host box
# ProxyCommand ferry pipe box 22
If the other side allows exactly one port, ferry pipe box alone reaches
it. With ferry up running here, pipes go through it.
Pairing does the same once and then swaps long-term keys. Later sessions between contacts authenticate with those keys instead of a code.
The rendezvous is not trusted with anything but availability. The threat model spells out what it and others can and cannot do.
Measured by the test harness (test/harness.sh): two machines, each behind
its own NAT, on one 16-core Linux box; netem shapes the path.
| path | ferry | one TCP stream |
|---|---|---|
| loopback, one machine | 2.0 Gbps | |
| LAN-like, through both NATs | 1.2–3.3 Gbps | |
| 40 ms round trip | 0.9–2.1 Gbps | 0.8 Gbps |
| 40 ms round trip, 0.5% loss | 0.64–0.69 Gbps | 0.4 Gbps |
Ranges are across runs; both ends share one machine's CPUs, so real machines on a real LAN do better. How the numbers were found and what was tuned: docs/throughput.md.
test/harness.sh builds two NATed homes, a rendezvous and a STUN server
inside a user namespace (no root needed) and checks: pairing; a wrong code
rejected on both sides; a direct path through both NATs; a plain "no
direct path" when UDP between them is blocked; a forwarded port through
strict NATs; no plaintext anywhere on the wire, rendezvous included; a
20 GiB file (FERRY_HARNESS_BIG=20480) killed part way over a 40 ms, lossy path and resumed with a
matching hash; and two MCP clients exchanging messages and a file.
Tested with 20 GiB files. Offsets are 64-bit and nothing holds a whole file in memory: during a 20 GiB transfer the sender used under 200 MB of RAM and the receiver about 130 MB.
MIT