Beacon: Open-Source Endpoint Telemetry for Local AI Agents

Beacon is Asymptote Lab's open-source endpoint agent that gives security and IT teams visibility into local AI agent activity. It hooks into agent harnesses like Claude Code, Codex CLI, OpenCode, Factory Droid, Claude Cowork, and Cursor, captures supported activity, and normalizes it into endpoint events stored as local JSONL or forwarded to SIEM pipelines.
Supported Agents and Outputs
The agent supports the following coding agent runtimes:
- Claude Code – via local OpenTelemetry configuration
- Codex CLI – via local OpenTelemetry configuration
- OpenCode – via Beacon hook adapter
- Factory Droid – via local OpenTelemetry configuration and optional hook adapter
- Claude Cowork – via admin-configured OpenTelemetry setup
- Cursor – via Beacon hook adapter
Outputs include local JSONL (default), Wazuh (localfile configuration with content pack), Elastic (Filebeat or Elastic Agent over JSONL), Splunk HEC (optional forwarding during install/repair), and customer-managed SIEM pipelines.
Architecture and Deployment
Collection, processing, and inspection remain local to the endpoint; forwarding is customer-controlled. Beacon filters out generic runtime metrics (Node.js event loop, V8 heap, CPU, memory) by default so agent prompts, tools, approvals, and file activity stay easy to inspect. To include runtime metrics, use beacon endpoint install --include-runtime-metrics or beacon endpoint repair --include-runtime-metrics.
For MDM deployment, Beacon supports Jamf Pro (macOS package, policy scripts, Extension Attributes) and Fleet (macOS package with user-context helpers).
Quickstart
Install via Homebrew:
brew tap asymptote-labs/tap
brew install beacon
For full setup paths, see the Quickstart docs. Security and IT teams should refer to the managed deployment guidance for rollout, validation, retention, and SIEM forwarding.
📖 Read the full source: HN AI Agents
👀 See Also

Reduce AI Coding Session Costs by 90% with Graph-Based Code Indexing
A developer built a local graph database that indexes a codebase using LLM-generated summaries, cutting Claude Code session costs from $6-10 to cents by avoiding redundant file re-reads.

Tracked 5 Biggest Claude Code SKILL.md Collections on GitHub — Sortable Table with Auto-Refresh
Built a sortable table of the top 5 skill-collection repos (totaling 125k stars) with star counts and skill counts, auto-refreshed by a /workflows:skill-collections command.

Two Free Claude Code Skills: Tutorial Generator and Prompt Fixer
Two new free Claude Code skills: create-tutorial generates code reading tutorials from your actual project files, and prompter rewrites typo-filled prompts into actionable instructions. Both are MIT licensed and install via GitHub.

AskFirst API adds human approval layer for AI agents
AskFirst is a REST API that lets AI agents pause for human approval before taking irreversible actions. It works with local models, hosted APIs, and any framework, providing email notifications, approve/deny options, and audit logs.