Repository navigation
A small release: clearer tool metadata, a consistent delete warning, and honest client reporting in HTTP mode.
Tool annotations are explicit on every tool
readOnlyHint, destructiveHint, and openWorldHint are now all set explicitly on all three action tiers and on list_endpoints and get_schema. destructiveHint=False on reads, writes, and the two discovery tools was previously left unset and therefore implicit, which clients are free to interpret as the spec default rather than as a deliberate claim:
| Tier | readOnlyHint |
destructiveHint |
openWorldHint |
|---|---|---|---|
read, list_endpoints, get_schema |
True |
False |
False |
write |
False |
False |
False |
delete |
False |
True |
False |
The delete warning matches the write warning
The delete tier's advisory text was DESTRUCTIVE - confirm with user before calling. while writes used a louder, fully capitalized form. Delete now reads WARNING. DESTRUCTIVE OPERATION. CONFIRM WITH USER BEFORE CALLING., so the more dangerous tier is not the more softly worded one.
As before, these warnings are advisory instructions to the client model. The server does not enforce confirmation.
streamable-http reports the real User-Agent
_resolve_ua_client_slug() matched the incoming User-Agent against a hardcoded list of known clients (claude, chatgpt, openai, cursor, codex, gemini) by substring, collapsing everything else into whatever happened to match first and erasing unknown clients. HTTP mode now passes the raw User-Agent header through unchanged. stdio still sanitizes clientInfo.name into a slug, which is unchanged.
Also in this release
- Every client configuration snippet in the README now shows the full recommended credential denylist rather than a single
system-keys:readtoken, so a copy-paste install gets the whole set.