verdande could not be added to Claude at all. The documented setup —
POST /api/v1/mcp with an Authorization header — cannot be configured from
Claude's custom-connector dialog, which asks for a name, a URL and OAuth
client credentials and has no field for a bearer token. The instructions
described something the interface does not offer.
Worse, the obvious guess failed silently in the most misleading way
available: /mcp was not a route, so the SPA fallback answered 200 with a
page of HTML. The connector reported a successful connection and then
failed to parse the response, which reads as a broken client rather than
a missing endpoint.
So /mcp exists now and takes its credential in the query string, exactly
as the calendar feed does and for the same reason: a client that cannot
be told to send a header means the address is the credential. The
request logger records paths without their query strings, so the key does
not reach the log.
It is deliberately outside /api/v1 and accepts the key and nothing else.
Reading a session cookie here would make it a state-changing POST outside
the CSRF check — a cross-site request that acts as whoever is signed in —
and there is a test that fails if that ever becomes true.
The API path is unchanged for anything that can send a header. The tokens
page now shows the connector address next to the token it belongs to,
since that is the one moment the token exists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gmail now connects. Authorisation-code flow with PKCE, the verifier kept
server-side rather than in a cookie — the callback is a top-level
navigation from Google, and a SameSite=Lax cookie is the kind of thing
that works until a browser tightens its defaults. State is checked, the
attempt expires after ten minutes, and the spent verifier is deleted
rather than left at rest.
Scope is gmail.readonly and the sync is one-way on purpose. Unstarring a
message does not delete the task; completing the task does not touch the
mailbox. Two-way would mean deciding what happens when both sides
changed, and the only honest answer is a sync model far larger than the
feature. The query carries newer_than:30d so connecting an account with
a decade of starred mail does not create a decade of tasks.
The update check is off unless the operator turns it on, asks GitHub at
most every six hours, and sends nothing about the instance — no
identifier, no version, no counts. A development build never reports an
update: somebody running their own tree is ahead of every release, and
telling them otherwise trains them to ignore the notice. Only
administrators see it, because only they can act on it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The MCP server speaks JSON-RPC over Streamable HTTP with the six tools
the brief names, authenticated by a personal API token — so a connector
reaches exactly what its owner reaches and there is no second identity
to reason about. A failed tool call returns a tool error, not a protocol
error: a model can recover from the first and can do nothing with the
second.
CalDAV serves VTODO, two-way, so Apple Reminders and Thunderbird can
read, create and complete tasks. Clients authenticate with Basic using
an API token as the password, never the account password — a client
stores what it is given, and what it is given should not be able to sign
in to the app. Completed tasks stay in the collection: a client that
stops seeing one treats it as deleted and drops its own record of having
done it.
The AI layer is one interface with adapters for Anthropic, OpenAI,
Google and any OpenAI-compatible endpoint. That last one is why the
abstraction exists: an integration that only spoke to a hosted API would
be useless to somebody running their own models. Every feature degrades
to "off" when nothing is configured, and the stored API key is never
returned to the settings page.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
