SECURITY
Agents you can audit
Switchyard is built on three commitments: agents act only with permission, every request is traced, and your data stays where it lives.
PERMISSION MODEL
Scoped by default
Every source connects with its own credential, and every key carries explicit scopes. Read and action permissions are separate grants. An agent holds exactly the access you gave it, nothing more, and a revoked scope stops working on the next request. Permission-aware execution means the check happens at run time, on every call, not once at setup.
AUDIT LOGS
Complete request tracing
Every request writes a full trace: caller, query, sources consulted, context returned, actions executed, permission checks, and latency. Traces are immutable, searchable, and exportable to your own logging stack. When a review comes, the evidence is already written.
DATA MOVEMENT
Nothing stored by default
Switchyard queries your systems where they live and returns context on demand. It does not copy your databases, mirror your documents, or build a shadow warehouse. Results pass through, get structured, and are delivered to your agent. What persists is the trace, not the data.
RETENTION
You set the clock
Trace retention is 7 days on Free and 90 days on Team. Enterprise sets custom retention, including continuous export to your own storage with zero retention on our side. Deleting a project deletes its traces.
Cloud or self-hosted, same API
Cloud
Run Switchyard in our cloud. Both deployments run the same API and SDKs, so nothing about your integration changes if you move.
Self-hosted
Run Switchyard inside your own network on Enterprise. Self-hosted means requests, traces, and credentials never leave your infrastructure.
Built for your audit, honestly
We build audit-ready controls: scoped access, permission-aware execution, immutable request traces, and exportable evidence. We document our practices plainly and support your security review with real answers, not badges. Ask us hard questions. We are ready for them.
Bring your security team
Read the full security documentation, or bring your security team to a call.
Talk to security