How it works
1
Connect
Point the OpenAI or Anthropic client your agent already uses at Ambral: a base URL, an Ambral key, and a few
request settings that name the task and the session. Your prompts, your tools and the calls your code makes
stay as they are. Requests are answered by the model you choose.
2
Record
Each session your agent runs is recorded as a trace: every model call it made, with its tool calls,
tokens and latency.
3
Evaluate
You build evals from the recorded sessions that matter, and say what a good reply looks like. Ambral replays
those sessions to each candidate model and scores its replies beside the ones production gave. The results
show the sessions where a model falls short, and what each model would cost to serve.
4
Switch
When a candidate holds up, you promote it. Ambral serves it within about a minute, your agent’s code doesn’t
change, and the same command takes you back. Nothing switches until you say so.
Two kinds of eval
A replay run scores candidates on your recorded sessions, one reply at a time. It needs no environment, no tools and no test suite, so it is where to start. An offline run is your own eval suite’s: the suite runs whole sessions on a candidate and reports the scores to Ambral. Replay runs show where to look. Offline runs check whole sessions before a candidate serves your users. Most agents should have both.How you use it
You drive Ambral with theambral CLI, and most people let their coding agent do it. The CLI installs a skill
into your project that teaches Claude Code, Codex, Cursor and other agents how to connect your app, build evals
from its recorded sessions, run them and read the results.
Where to go next
Quickstart
Install the skill, ask your agent to connect your app, and see your first trace.