Sci Alpha — Install Guide
Alpha v0.1.0 · macOS · Apple Silicon · Invite only
What you're getting
Sci is a local proxy that sits between Claude Code and Anthropic. Before any request leaves your machine, it substitutes your real name, email, and other PII with stable placeholder tokens, then restores them in the response. It also injects relevant memory context — facts about you, your projects, and past decisions — into every session via an MCP server connected to a local store. The alpha proves the core pipeline: anonymization, memory injection, and deanonymization all working end-to-end with Claude Code. What's not included yet: auto-update, a GUI, multi-provider routing, and cloud storage sync.
Prerequisites
- macOS 13 or later, Apple Silicon (M1 or newer) — Intel Macs are not supported in the alpha
- Claude Code installed and signed in (
claude --versionshould succeed) - An invite link from Casey
No API key required — Sci uses the Claude Code login you already have, including Claude Pro or Max via OAuth. Storage is local SQLite — no Docker or Postgres either.
Install
Run the installer in your terminal:
curl -fsSL https://raw.githubusercontent.com/sovereign-cognitive-identity/sci/main/scripts/install.sh | bashThe installer asks for
sudofor two steps: trusting the CA certificate in your System Keychain and writing the launchd services.
The installer will:
- Download the Sci binary for Apple Silicon and install it to
~/.sci/bin/ - Add
~/.sci/binto yourPATHin~/.zshrc - Generate a local CA certificate and trust it in your macOS System Keychain
- Register the
sciMCP server in~/.claude.json - Write
HTTPS_PROXYandNODE_EXTRA_CA_CERTSinto Claude Code'ssettings.json - Install and start the launchd services —
dev.sci.helper(proxy, :3001) andcom.sci.agent(:8080)
When it finishes, you'll see a status summary.
If the installer offers to store an Anthropic API key, you can skip it (press Enter) — your existing Claude Code login already works through the proxy. A key is only needed if you prefer key-based auth over OAuth.
First-run setup
The installer completes setup for you — there is no separate wizard to run. If you ever need to re-enter your API key or re-register the MCP server, run:
sci --setupVerify it's working
Open a new terminal (so the PATH change takes effect), then:
sci statusExpected output:
ok: true
episodic: 0
semantic: 0
identity: 0
embeddings: 0
queue (pending): 0Counts are zero on a fresh install. That's expected — they grow as you use Claude Code.
Then run a live privacy check:
sci verifyThis sends a real request through the proxy and prints a before/after comparison showing what left your machine. The check must PASS before you trust the system with real conversations.
Sample output:
What you sent:
"My name is Casey Zandbergen and my email is casey@example.com."
What reached Anthropic:
"My name is [PERSON_1] and my email is [EMAIL_1]."
Anonymization:
PERSON_1 → Casey Zandbergen
EMAIL_1 → casey@example.com
PASS — real name did not appear in the outbound request.Your first session
- Open a new terminal (PATH and proxy env must be inherited)
- Start Claude Code in any project:
claude - Confirm the proxy env is active — ask Claude to run
echo $HTTPS_PROXY. It should printhttp://127.0.0.1:3001. - Use Claude Code normally. Sci intercepts each request transparently.
- Start a second Claude Code session in a different directory. This is where memory injection becomes visible — Sci will include context from the first session without you doing anything.
See Using Sci for day-to-day usage.
Confirming memory is flowing
Sci exposes memory to Claude Code through the sci MCP server. Ask Claude directly:
"What do you know about me and my active projects?"
Claude will call memory_identity and memory_recall and summarize what it finds.
To seed memory immediately from your Claude conversation history:
# Export from: claude.ai/settings/account → Privacy → Export Data
sci import --claude ~/Downloads/conversations.jsonUninstall
# Stop and remove the background services
launchctl unload ~/Library/LaunchAgents/dev.sci.helper.plist
launchctl unload ~/Library/LaunchAgents/com.sci.agent.plist
rm ~/Library/LaunchAgents/dev.sci.helper.plist ~/Library/LaunchAgents/com.sci.agent.plist
# Remove MCP registration
claude mcp remove sci
# Remove config, binary, and data (destructive — erases all stored memories)
rm -rf ~/.sci
# Manually remove the PATH line from ~/.zshrc and the proxy env block
# (HTTPS_PROXY / NODE_EXTRA_CA_CERTS) from Claude Code's settings.json
# Remove CA from Keychain (optional):
# Keychain Access → search "Sci" → delete the certificateCommon issues
| Symptom | Cause | Fix |
|---|---|---|
sci: command not found | PATH not updated | Open a new terminal or run source ~/.zshrc |
Connection refused on port 3001 | Proxy not running | launchctl load ~/Library/LaunchAgents/dev.sci.helper.plist, then check ~/Library/Logs/sci-helper.log |
| First request hangs 30–120 seconds | Embedding model downloading on first run (~110 MB) | Wait — tail -f ~/Library/Logs/sci-helper.log to watch progress |
curl: (60) SSL certificate problem | CA not trusted in Keychain | Re-run the installer, or trust ~/.sci/ca.crt in Keychain Access manually |
| 401 from Anthropic | Claude Code isn't signed in (or a stale key is set) | Run claude and sign in; if you use an API key, sci --setup to re-enter it |
| Claude Code bypasses proxy | HTTPS_PROXY not set in Claude Code's environment | Launch Claude Code from a terminal, not the Dock/Spotlight |
See INSTALL.md for detailed diagnosis steps.
Feedback
This is alpha software. Things will break. Please report what you find:
- GitHub Issues: github.com/sovereign-cognitive-identity/sci/issues (preferred)
- Email: casey.zandbergen@gmail.com
When reporting a bug, include the output of sci status and the relevant lines from ~/Library/Logs/sci-helper.log.