Open source. Self-hosted. Apple Silicon, macOS 26 or later.

See what your AI agent actually did.

Agent transcripts show what the agent reported. RawTrail independently records process execution, file changes and network activity at the macOS level, then connects that activity to AI agent sessions.

View on GitHub Read the docs

3tool calls the agent reported
109process executions the Mac recorded
One Claude Code session. The agent declared 3 tool calls. The Mac recorded 109 process executions, and fewer than a quarter of them descend from those calls.

Why RawTrail

How it works

  1. Record. A macOS system extension and network extension watch process, file and network activity and tie it to the AI agent session behind it.
  2. Collect. A small agent on each Mac sends the events to your RawTrail server, and queues them if the server is unreachable.
  3. Review. Open the web UI to see each agent session as a process tree, with alerts where something looks wrong.

Built on a macOS EDR

RawTrail is built on a macOS endpoint detection and response system. Its behavioral detections cover credential access, persistence and command and control, mapped to MITRE ATT&CK, and application control can block execution at the operating system level.

Get started

Try the server and web UI without a Mac. The demo runs in Docker with sample hosts and detections. It is for evaluation only, so do not expose it to the internet.

git clone https://github.com/RawTrail/RawTrail.git && cd RawTrail && docker compose -f docker-compose.demo.yml up

Then open https://localhost:8088/ui/.

Deploy it. Run the server on one Linux VM, then push the signed agent package to your Macs with any MDM. See the single-VM quickstart and the MDM guide.

Status. Pre-1.0 and under active development. RawTrail records and attributes agent activity today. It does not yet enforce policy per agent session. Session tracking covers Claude Code, and other agents are still recorded as ordinary processes. Releases are signed and ship with build provenance.