Agents are dope and powerful. They are also non-deterministic. Frontier models do the right thing most of the time and still fail in ways nobody predicted. Just this month a frontier lab's eval agent broke containment and compromised a model host on the way out. If agents go off the rails inside the labs that build them, what chance do the rest of us have of controlling them 100% of the time? Guardrails prevent. Provenance explains. When an agent does something surprising, the answer to why currently lives in scattered client logs or nowhere.
Scanduit is a proxy in front of your LLM traffic. Point
base_url at it and every agent you run is on the record. Every
tool call out, every result back, every token, which model actually made
each decision and which identity made the call are all captured and recorded.
All of it parsed straight from the stream. Zero agent cooperation required.
Reasoning is recorded when the provider exposes it. When they don't provide, which is the more common case, we reconstruct it from the tool call trajectory, which we see completely. The transcript is the record. Every session lands in a ledger built for one question: why did that run do what it did. Not another cost dashboard.
Adoption is one line.
- base_url = "https://api.anthropic.com" + base_url = "https://proxy.scanduit.com"
Override base_url and everything gets recorded. That includes
your agents, third party agents, even the dev agents running on laptops
nobody audits. The proxy holds the provider keys. Agents hold Scanduit keys.
Fail closed: no record, no session. A gap in the ledger is a first-class event, never
silence.
An agent is whatever authenticates with a Scanduit key. Identity attaches at the key, and that is where limits will attach too: caps on tools, models, tokens and egress, enforced at the same choke point that records them. Recording ships first. You can't limit what you can't see.
Building now. Leave an email, get the updates that matter.
SCANDUIT · © 2026