Chain of thoughts.
We calculate base probabilities for possible causes, so you can understand your agent’s behavior and what may be driving each step.
Explore chain of thoughts →STORAGE LAYER FOR YOUR AI EVIDENCE
We don’t. That’s why we built a system to observe everything. Every file access, every network call. Any command that steps outside the boundaries you define is reported back to you with stored evidence.
ONE SESSION. EVERY CONNECTION.
Every read, write, and network call.
Connected to the evidence.
Hover to rotate · drag on touch · arrow keys when focused Illustrative connections stay attached as the sandbox turns.
ONE CLI / GET STARTED
Give your AI a sandbox.
Claude Code and Codex come preinstalled in every sandbox. Your agent’s workspace is ready when you are.
raytrace sandbox create --localConnect to your sandbox, sign in to Claude Code or Codex, and run your agent harness. RayTrace observes as it works.
Check the dashboard, investigate what happened, and store the evidence locally. Run out of space? Upload it to RayTrace Cloud.
MAKE SENSE OF YOUR AGENT
We calculate base probabilities for possible causes, so you can understand your agent’s behavior and what may be driving each step.
Explore chain of thoughts →Change what your agent saw. Fork the run from that step, continue in a separate sandbox, and compare what it does differently. See which actions, file changes, and results follow from your edit.
Explore a forked run →Math.floor(19.6)19Math.round(19.6)20Search your stored evidence across runs. Find failed commands, compare reported success with observed results, and pinpoint the step where your agent went off track.
Explore recorded evidence →Expected: 20 · Received: 19 · exit: 1OBSERVED FAILUREThe agent reported success. The recorded test failed.
The initial scan highlights sensitive and important files in your project. Choose what to restrict or watch. If your agent reads a restricted file like .env, RayTrace records it and raises an alert—with an email when notifications are enabled.
Explore the initial scan →.envSENSITIVERestrictedmoney.jsWatchedREADME.mdStandard Read .env.env accessed · evidence recorded
Set custom restrictions on agent file actions. See which files it reads and which destinations it tries to reach, with the evidence kept in one place.
Explore restrictions and network activity →.envRESTRICT CHANGESmoney.jsWATCH ACCESSBoundaries you define.
Actions you can inspect.
read money.jsOBSERVEDconnect anthropic.comATTEMPT