Your people have AI on their laptops. Beam gives you a per-machine record of what that AI did — commands, file changes, tool and MCP calls, endpoints — stored locally and readable when someone asks.
It sits beside your existing endpoint tooling, and it observes rather than enforces.
Everything below is verifiable on one machine before you touch a second one.
Footprint
The collector binds to 127.0.0.1:4319 and writes locally. No kernel extension, no browser extension, no OS surveillance to review.
Deployment
A hook merged into settings plus a local service. Scriptable, but there is no signed Jamf, Kandji or Intune package today.
Coexistence
Beam watches the layer endpoint tooling cannot see: tool calls, MCP servers and skill files. It intercepts nothing at the OS level.
Portability
OTLP/HTTP JSON in, redacted NDJSON out. Nothing is forwarded automatically, so the data path is only what you build.
Supply chain
Eleven heuristic patterns plus a version-pin check, run before someone connects a file to production credentials.
Support
Flagged actions keep the triggering event as evidence, so a ticket closes with a record instead of a recollection.
By 2028 (Gartner prediction)
Gartner also puts spend on securing AI near $4.8 billion in 2027. Endpoint owners inherit most of that question.
July 2025
A prompt injected through a pull request reached v1.84.0 of the Amazon Q extension. Software-update review does not cover instruction-layer risk.
September 2025
postmark-mcp v1.0.16 added a silent BCC to an attacker address, on 1,643 total downloads. People install these without a procurement step.
01
Run the collector, pair the app with its printed token, import a sample record. The token stays in browser memory, not in a URL.
02
Hooks are merged deliberately, so nothing changes on a laptop without an action someone took on purpose.
03
Use two weeks of real activity to write policy grounded in what your AI does.
Beam is built so you can say precisely what it does — and what it cannot do.
Not as a packaged deployment today. Pairing is designed for local single-machine use, and fleet management, SSO and database-backed teams are documented as future work. Setup is scriptable per machine — enough for a pilot, not a managed rollout.
Only if you describe it badly. Beam records what the AI did, not what the person did: no keystroke logging, no screen capture, no browser extension. Data stays in local files on their own machine, and nothing is forwarded automatically.
It should not. Beam is a userspace process on loopback that receives events a hook sends it. No kernel extension, no OS-level interception — and it covers the AI layer endpoint tooling has no visibility into.
Work carries on. The hook sends a bounded request and emits no allow or deny response, so nothing waits on a verdict. The cost of a failure is a gap in the record, not a blocked person.
Find out what AI is doing on your endpoints before you write the standard — with a record you can show the person it came from.