A local proxy that wraps your MCP servers and checks each tool call against policy and live state before it runs - allow, block, or request a refresh, with a reason the agent can act on.
-
Updated
Jul 12, 2026
A local proxy that wraps your MCP servers and checks each tool call against policy and live state before it runs - allow, block, or request a refresh, with a reason the agent can act on.
Hermetic sandbox executor that returns typed, delta-aware feedback to coding agents (Rust). Parsed compiler/sanitizer/test verdicts + run-to-run deltas instead of raw stdout.
Send signals about your AI agent's work to Glassray: what people do with it, and what the agent says about it. One line for MCP servers, a setup skill for coding agents, and a check for your tests.
MCP server that lets AI agents log friction to .vents/ markdown files in your repo. Open-source implementation of Lovable's vent-tool pattern.
Claude skill with a custom widget for per-section feedback on drafts. Target multiple sections at one, provide guidance, and get a re-draft
Pre-registered replay of 80 failing states: a vague verifier error is fixed at the next retry 12.5% of the time, expected-vs-actual types 45%. Includes verdict-lint, an offline audit of agent verifier loops. Fictional platform, synthetic data, Haiku 4.5, single step. Playbook memory: not significant.
Observe every move. Diagnose every waste. Feed every lesson back.
To associate your repository with the agent-feedback topic, visit your repo's landing page and select "manage topics."