This whole stack has one single point of failure: an account. Lose it and everything local still exists but cannot say a word. Causes, backups, and legitimate rerouting — the backup list is the only part that needs action today.
In this pagecommon causes at the terms level, a backup list, legitimate rerouting. Not in this pageevading detection, multi-accounting, working around risk controls. The reason is in §5 and it is not boilerplate.
Lower your expectations first: most account problems are not "caught doing something bad". They are three lines drawn closer than people picture.
| Kind | Typical | Why it catches people |
|---|---|---|
| Shared access | Handing account access to someone else, or reselling it | Nearly every terms document forbids sharing and resale — but "sharing with a colleague" feels ordinary |
| Automation beyond the licence | High-frequency unattended calls beyond what that plan permits | Subscription plans and API plans often permit different things — people assume a subscription covers automation |
| Payment and identity | Failed charges, regional restrictions, duplicate signups from one identity | The unluckiest kind: not behaviour, just automated billing and compliance decisions |
Every vendor words it differently and all of them revise. Copying clauses into this page means they start going stale immediately, and a reader following stale terms is worse off than one following none.
Read the terms of the service you actually use, looking for three words: share, automated, resell. Those three roughly bound whether your usage sits inside the licence.
Losing an account is survivable. Losing six months of accumulated state alongside it is not. These live separately from the account, and most people have never backed them up:
| Back up | Where | Cost of losing it |
|---|---|---|
| Config | Each tool's config directory (~/.claude/, ~/.codex/, ~/.hermes/…) | Permission rules and provider setup, rebuilt from scratch |
| Project memory | CLAUDE.md / AGENTS.md in each repo | Usually in git — the one category that is safe by default |
| Custom commands / skills | Tool user directory | Hurts more the longer you have run; rarely version-controlled |
| Tool wiring (MCP etc.) | Server list in config | Reconnecting means re-reading a lot of docs |
| Session history | The tool's local session store | "How did I solve this last time" — gone |
| Key inventory | ⚠ List what exists and where to reissue. Never the values | Knowing which credentials to re-obtain |
Fold those config directories into whatever backup you already run — time machine, sync folder,
private repo — excluding the file that holds secrets.
Most tools deliberately keep keys in a separate .env-style file precisely so the rest is safe to copy.
Confirm which file that is before you back anything up. The Hermes page
gives a concrete case: settings in config.yaml, keys in .env, split on purpose.
"Rerouting" has exactly one honest meaning: moving to a provider you legitimately have access to. Not a new identity at the same one.
The good news is that it is technically cheap, because these clients speak a protocol, not a vendor — switching is an endpoint and a key. That is all of the next page.
| Change | Cost |
|---|---|
| Endpoint + key | One config edit |
| Model name | One config edit, but re-learn the feel — models answer the same prompt differently |
| Protocol differences | This is the real cost. Some endpoints support only one request shape and something breaks. Page 7 §4 |
| Do | Why |
|---|---|
| Keep the endpoint a config value, never hard-coded | Switching becomes one line, not a search across a codebase |
| Keep one alternate path you could switch to now | You do not need to use it. You need to have switched to it once and seen it work |
| Actually switch, occasionally | An unrehearsed fallback is not a fallback. Old ops truism, still true here |
| Do not hang critical work on a single account | See page 21 — plan for a long-running service going dark |
Plainly: not a moral pose. Two pieces of arithmetic.
One. That content has negative expected value for the reader. Evasion techniques have a half-life measured in weeks; a web page gets read for years. Following a stale evasion guide is more dangerous than doing nothing.
Two. The very next page covers relay endpoints — a completely legitimate engineering technique, and a different thing entirely from wearing a new mask. If this page taught the latter, the next page gets read as the same category. That trade is far too expensive.
~/.hermes/config.yaml and ~/.hermes/.env,
checked 2026-08-21). Confirm which file holds secrets in your own tools first.