Skip to main content
Belvedir traces your LLM agents, segments each run into the tasks the agent performed, and uses those tasks to route calls, fine-tune models, and propose changes to your agent’s code. Coming from Fractal? See Migrating from Fractal.

​
Set up in this order

1

Get traces flowing

The Quickstart (Node) or the Python quickstart: install, initialize, wrap runs in sessions.
2

Fit it to your stack

Guides for Next.js and Vercel, the AI SDK, custom agent loops, background workers, and agents without the SDK. Bring history over with Importing traces.
3

Route your calls through Belvedir

Two lines change and Belvedir executes your LLM calls. Start at Route your calls.
4

Improve the agent

Optimization loops propose changes to your agent’s prompts and code. LoRA finetuning trains a small model on your successful traffic. Benchmarks measure both.