Cerver sessions are the control layer for your AI stack. They can compare intelligence, move work between compute providers, spawn parallel runs, track cost, and keep the product interface stable while models and runtimes change underneath.
Provider secrets stay in Infisical. Cerver resolves them when a session needs them, then shows what ran without exposing the raw key.
Use OpenAI, Claude, Grok, Gemini, local models, or your own adapters for the same intent. Compare the answers and keep the strongest result for the task.
Your product talks to the Cerver session. The session can run work on Vercel, Cloudflare, E2B, local compute, or another provider without rebuilding the transcript or changing the product flow above it.
An agent can create five versions of the same task with five intelligence layers, compare the outputs, choose the best one, and keep the full trail.
Simple requests can use cheaper paths. Risky work can use stronger models and compute. The session keeps one record while the route changes underneath.
Tool calls fail. Tabs close. Providers change response shapes. Cerver keeps the session boundary clear so your product has a stable place to continue.
Every session can show which model ran, where it ran, what it cost, and why. That makes AI spend easier to explain and optimize.