Skip to main content
Belvedir is an observability and auto-improvement platform for LLM agents. Install the SDK, wrap your agent runs in sessions, and Belvedir segments the traces into tasks, gathers similar tasks into training sets, and runs improvement loops (harness evolution, prompt evolution, and LoRA fine-tuning) against your production traffic.

​
Where to start

Quickstart

Get traces flowing in under 5 minutes with the Node SDK.

Python SDK

Same concepts, context managers instead of callbacks.

How it works

The six-stage pipeline: instrument, collect, segment, gather, clean, optimize.

Agents without the SDK

Point any OpenTelemetry exporter at Belvedir: OpenClaw, Hermes, or your own.

​
What you can do

Trace

Every LLM and tool call from OpenAI, Anthropic, Cohere, Bedrock, Vertex AI, and more. Sessions segment into tasks; similar tasks gather into training sets.

Route

Send each call to the model that fits it, decided from your own task groups, and let the record of what worked pick the cheapest model that wins.

Optimize

Harness evolution and prompt evolution loops that feed on production traces and open PRs against your repo, plus LoRA finetuning on your successful traffic.

Benchmark

Run any agent harness in an isolated CPU or GPU sandbox against SWE-bench, τ-bench, GAIA, and more, or environments generated from your own training sets.

​
Supported runtimes

The Node and Python SDKs (both belvedir — npm and PyPI) both auto-instrument the major LLM providers. (The legacy @fractalresearch/loop and fractal-loop packages still work but are deprecated.) Local models via Ollama and hosted OpenAI-compatible endpoints (Together, Fireworks, Groq, OpenRouter, vLLM) work out of the box. See Cloud Inference.
Working with a coding agent? The whole integration guide lives at platform.belvedir.ai/agents.md in agent-readable markdown: both SDKs, outcome reporting, and the common pitfalls. It’s also published as a skill at github.com/Belvedir/skills. The dashboard’s Quick Start page has a Copy prompt for AI button that puts a ready-made instruction on your clipboard. Paste it into your coding agent and it downloads the skill into your codebase (as .claude/skills/belvedir/SKILL.md) and instruments the app for you.