Anthropic Cloaked Proxy — Development skill for Claude Code

Anthropic Cloaked Proxy for Hermes

Run Hermes Agent through a Claude Code OAuth fingerprint. Uses a two-proxy chain so personality, skills, and tools all work natively.

Hermes → Soul Proxy (:8319) → Cloaked Proxy (:8318) → Anthropic API
        injects SOUL.md          tool mapping (28→14)
        + skills list            CC fingerprint cloak
        + memory context         OAuth auth

Why two proxies

• Anthropic fingerprints the system prompt — must stay under ~150 chars
• Personality injection has to happen in the first user message
• Soul proxy handles persona + context injection
• Cloaked proxy handles tool mapping + fingerprint cloaking + OAuth headers

Install Claude Code

⚠️ Use the official install script, **not** npm:

curl -fsSL https://claude.ai/install.sh | bash

Then log in (browser flow):

claude login

Deploy proxies

Copy the three files to your Hermes host (they live in `~/proxies/` on the current install):

scp cloaked-proxy.py soul-proxy.py start_both.sh you@your-host:~/proxies/

Start both:

bash ~/proxies/start_both.sh

Point Hermes at the soul proxy:

hermes config set model.provider anthropic
hermes config set model.default claude-opus-5
hermes config set model.base_url "http://127.0.0.1:8319"
hermes config set model.api_key ""

`api_key` is intentionally empty — the cloaked proxy supplies OAuth credentials upstream, and a non-empty value here sends a conflicting `Authorization` header.

Test

# Personality check
hermes chat -q "Without using any tools or skills, what can you see are available from your system prompt."

# Tool execution check
hermes chat -q "ls ~ | head -3"

Run the test suite (needs the interpreter pytest is installed under; unset `PYTHONPATH` so a venv on the path doesn't shadow it):

cd ~/proxies && env -u PYTHONPATH python3 -m pytest tests/ -q      # 28 passed, 1 skipped
cd ~/proxies && env -u PYTHONPATH python