Show HN: Intercept and control AI agent actions before they happen (github.com)

Echos is an open‑source "firewall" for AI agents that intercepts actions before they execute, letting you enforce policy or require human approval for sensitive operations. Developers integrate an SDK (npm/pnpm) so agents emit structured actions (e.g., slack.post, http.request, fs.delete), which a local daemon evaluates against regex‑based policies and returns allow, block, or ask. If an action is marked "ask" the dashboard surfaces a consent modal and a human can approve once or grant a time‑limited token; agents can also request persistent authorizations programmatically. The demo and repo include ready role templates (research_assistant, customer_support, internal_notifier) and examples to get started quickly. For the AI/ML community this is significant because it provides real‑time, provenance‑aware control over agent behaviors—preventing unsafe side effects like unintended file deletion, rogue HTTP calls, or noisy notifications—while preserving developer ergonomics and offline/local control. Technical highlights: scope list (llm.chat, http.request, email.*, slack.*, calendar.*, fs.*), policy hot‑reload, live feed/timeline and audit logs, metrics dashboard, policy testing/validation tools, and token lifecycle management. Security defaults block unknown intents, ReDoS protection (100ms regex timeout), API timeouts (3–10s), JWT (HS256) auth, local‑only binding by default, and metadata‑only logging with NDJSON export for compliance. Licensed MIT.