The right model for every task. You stay in control.
Synapse does not bet on a single AI vendor. You set the model for each agent once and every block and flow inherits that choice: precision where it matters, savings where it does not compromise. Automatic model recommendation is on the roadmap.
Pick a task and see who takes it
The decision happens before the call, based on what the block declares: task nature, criticality, input format, and context size.
One application. Several models working together.
Routing isn't per application: it's per block. A single application can call four different models in one run.
What the router does beyond choosing
The part that picks the model is the most visible. The rest is what makes it work in production.
Automatic fallback
If a provider goes down, gets slow, or returns an error, the call moves to the next compatible model. The application doesn't stop because a vendor had an incident.
Cost cap per squad
Each squad has a monthly budget. As it nears the limit, the router steps down to cheaper models on light tasks and alerts the owner — instead of silently blowing up the bill.
Context caching
Policy excerpts, catalog entries, and system prompts that repeat on every call aren't paid for again on every run.
Per-call log
Model used, input and output tokens, latency, cost, and the block that originated the call. It's what lets you audit a decision months later.
Swap without rewriting
The block declares the task, not the vendor. When a new model joins the platform, it becomes a candidate for compatible tasks — without touching the applications already running.
Data that doesn't leak
No call authorizes training on your data. For sensitive content, the router can be restricted to open models hosted in your environment.
The router picks the brain. The base decides what it knows.
Want to see the routing with your own tasks?
Bring three real tasks from your operation. In the technical demo we show which model each one would call, the estimated cost per run, and what the audit log looks like.