AgentBeam

AgentBeam is a local-first control layer for AI coding agents that records their shell commands, file edits, tool calls, and network activity, flags risky actions with evidence, and scans MCP configs and SKILL.md files for tool poisoning and credential exfiltration before an agent trusts them.

AgentBeam is a local-first security and observability layer for AI coding agents. As developers hand more work to terminal and IDE assistants that can run shell commands and edit files, it becomes hard to see what those agents actually did and whether any of it was risky. AgentBeam attaches to the coding agents already running on a machine, with no change to how a team prompts them, and captures shell commands, file edits, tool calls, and network activity into a single local record as the agent works. On top of that record, AgentBeam surfaces the actions that matter. Sensitive-data access, credential exposure, destructive commands, and permission changes are flagged along with the evidence behind them, so a reviewer can see not just that something happened but why it was risky. A distinct strength is Model Context Protocol scanning: before an agent connects to a new MCP server, AgentBeam checks its configuration and its SKILL.md files against heuristic patterns for tool poisoning, unpinned versions that enable rug pulls, and credential-exfiltration phrasing, catching hidden instructions before the agent treats them as authoritative guidance. The design is local-first. Activity is recorded to files on the machine where the agent runs, and nothing leaves the device unless an export is explicitly configured, which suits developers and security teams who want visibility without shipping every interaction to a third party. AgentBeam is distributed as an open-source CLI that can be initialized to trace agent activity automatically, and it produces audit-ready evidence. It fits individual developers who want a safety net on their own machine, security teams that need detection and evidence across every agent in use, and IT teams rolling out AI coding tools across developer machines who need fleet-wide visibility.