Skip to main content
Claude Code supplies the interaction loop and tools; the selected model supplies the predictions. They are different parts of the agent. Synth connects them through your Synth identity, so a project does not need its own provider key.

​
Launch an interactive session

Install Claude Code separately, then install Synth and sign in.
The current Synth command handles its own help without beginning a model session. Native Claude arguments after -- belong to Claude Code; their availability depends on the installed native version. A model offered to one account or connection need not be offered to another.
Interactive Claude Code can use tools on your machine, subject to the native tool’s permissions. Launching it through Synth does not put your working directory into a cloud sandbox.

​
Move from interactive work to repeatable work

Use an agent profile to pin the model, harness version, files, tools and execution recipe for managed tasks. Do not assume an interactive session’s local settings or permissions automatically transfer to that profile. A profile can run locally for debugging, but local execution is also not a sandbox. Use managed execution when the work should run in a task’s declared environment.

​
Inspect what happened

Session evidence explains streams, interactions and exports. Do not confuse an interactive session with a scored evaluation: to measure or train an agent, define a task and its assessment.