LangSmith + Memtro

Only when you ask, and only what it has been granted: add feedback (a key with a score, value or comment) to a LangSmith run; create a new LangSmith dataset; add an example to a LangSmith dataset; edit a LangSmith dataset example's inputs, reference outputs, metadata or split; rename or re-describe a LangSmith dataset; change the score, value, comment or correction of LangSmith feedback; create a LangSmith annotation queue for people to review runs; edit a LangSmith annotation queue's name, rubric or reviewer settings; put named runs into a LangSmith annotation queue for review; create a LangSmith tracing project; rename or re-describe a LangSmith tracing project; delete a LangSmith dataset example permanently; delete a LangSmith feedback record permanently; delete a LangSmith dataset and all its examples permanently; delete a LangSmith tracing project and all its runs permanently; delete named traces from a LangSmith project permanently. Every other tool reads, searches or lists. The Connections page labels the tools that act, an administrator grants each one by name to the teams that may use it, and a personal connection acts only after its owner allows it.

A personal connection serves only your own requests. A shared connection serves the organisation: everyone, or only the teams an admin chooses. Admins can always add shared connections; other members need the permission.

Credentials are encrypted at rest. Tool results are cached briefly so agent loops do not repeat identical calls, and the facts the model extracts are saved to your organisation's memory with provenance (system, record and link) so they can be checked and superseded.

API key. In LangSmith, open Settings → API Keys → Create API Key and choose a Service Key with a role that can read.

Yes. Turn events on in the connection's Events tab on the Connections page, then add the address it shows as a webhook in LangSmith, choosing the same kinds; the tab shows LangSmith's own steps. A job set to start "When an event arrives" then runs once for each matching event, with the event as its input. The run treats the event as outside data, never as instructions, and anything the run remembers is kept as unverified.