Archcore Send

Hand off an AI coding session — encrypted, one-time, zero-knowledge.

Package the working context of one session — goal, state, errors, diffs — encrypt it locally with age, upload only ciphertext, and get a one-time link. The other side picks up exactly where you left off — on another machine, another model, or with a teammate. The server only ever sees ciphertext.

You're three hours into debugging with one agent. Now you want to continue on another machine, with another model, or with a teammate — without re-explaining everything or pasting secrets into a chat box. That's what /send is for.

How it works

One round trip. You run /send, hand over the link, the other side loads it.

# you — package the current session into an encrypted one-time link
/send                 →  🔗 https://send.archcore.ai/s/snd_01J…#agekey=AGE-SECRET-KEY-…

# them — load the link, resume with full context
/send --load <link>   →  ✅ Loaded "Auth debugging handoff" — compact-first.

The decryption key lives in the link's #agekey=… fragment and is parsed locally by the client — it never touches the network. Load is compact-first: a tight, high-signal context (~8k tokens) by default; big diffs and logs stay as optional, lazy-loaded details.

🔐 The full link is a secret. Anyone holding the complete URL — including the part after # — can decrypt the context. Share it like a password, prefer one-time links, and don't paste it into untrusted remote tools.

Install the skill

The thing you install is the skill — one self-contained folder. No build step, no binaries, fully auditable. Two steps:

1. Copy the folder into your agent's skills directory.

cp -R skill/send ~/.claude/skills/send     # Claude Code
cp -R skill/send .agents/skills/send       # Cursor · Codex · Gemini · Zed · opencode · …

SKILL.md is the open Agent Skills standard, so the same folder installs unmodified across Claude Code, Cursor, GitHub Copilot, Codex CLI, Gemini CLI, Zed, opencode, Windsurf, Cline, and more — only the destination dir differs. The full per-agent matrix is in skill/send/INSTALL.md.

2. Check it works.

By default the skill talks to the free public instance https://send.archcore.ai — zero config. It's zero-knowledge (the server only ever stores ciphertext, never your context or keys), so it's safe to use as-is:

/send --doctor   # checks age / curl / gzip + server reachability

That's it. Run /send. Prefer your own server? See Self-hosted.

No skill support in your tool? Run skill/send/scripts/send.sh directly — the script is the whole product.

Why it's different

• 🔒 Zero-knowledge server. It stores only ciphertext + operational metadata — no plaintext, no titles, no key. (It can't redact what it can't read; the trust boundary is small and auditable on purpose.)
• 🗝️ The key lives in the link. End-to-end encryption with age; the decryption key rides in the URL fragment and never reaches the server.
• 🎯 Compact-first load. A tight, high-signal context by default; large diffs and logs stay as optional, lazy-loaded details.
• 🧹 No binaries, fully auditable. Just SKILL.md + a shell/PowerShell script orchestrating age, gzip, and curl. Read every line before you run it.

Quick start

/send                 # package current context → encrypted one-time link
/send --load <url>    # load a received link into this session
/send --inspect       # preview what WOULD be sent — uploads nothing
/send --doctor        # check age/curl/gzip + server reachability

Useful flags: --ttl 24h (max 7d) · --one-time (default) · --yes · --allow-secrets · --include-large · --server <url>.

See skill/send/README.md for the full client guide.

Self-hosted

The public send.archcore.ai is the zero-config default, but self-hosting is a first-class path. Two parts: run the server, then point the client at it.

1. Run the server. Any VPS with a public IP and Docker — the design is provider- and storage-agnostic. From the repo:

cp deploy/.env.example deploy/.env    # set SEND_DOMAIN + SEND_PUBLIC_URL
cd deploy && docker compose up -d --build

Caddy fronts sendd with automatic TLS. Full guide: deploy/README.md and .archcore/guides/self-host-deploy.guide.md.

2. Point the client at it. Override the built-in default with SEND_SERVER_URL (persistent) or --server (per call):

export SEND_SERVER_URL="https://send.example.com"   # or: /send --server https://send.example.com
/send --doctor                                       # confirm it's reachable

Repository layout

This is a monorepo with two independently-packaged deliverables:

Part	What	Install
skill/send/	The /send Agent Skill (client) — SKILL.md + shell/PowerShell scripts. No binaries.	copy the folder into your agent's skills dir
server/	The Send server (Go, zero-knowledge ciphertext rendezvous) — single static binary / container.	self-host via deploy/

The skill is self-contained: copying skill/send/ carries no server or Go files.

Development

make lint          # shellcheck the skill scripts
make test          # bats unit + e2e (e2e uses a python mock server)
make server-test   # go test ./... -race
make test-all      # skill + server + e2e against the real Go server

Architecture & decisions

All architecture, decisions, specs and rules live in .archcore/ (Git-native context). Start with architecture-overview.doc.md and the specs under .archcore/specs/.

License

Apache 2.0.

---

Source: github.com/archcore-ai/send · Built by @ivklgn.