Agent operating suite
Tools
Tools for building, testing, operating, and supervising AI agents, built alongside this failure registry. Everything below is live.
Agents & operations
Support inbox operator that auto-resolves within policy and escalates the rest with a proposed action.
One-tap human-in-the-loop approval cockpit for agent escalations, on web and mobile.
Support triage as a LangGraph state machine: typed state, checkpointing, and a human interrupt for risky actions.
Computer-use agent for tools with no API: it drives a real browser and refuses any write until you confirm it.
Scoped MCP server that exposes company tools to agents with per-client permission boundaries and an audit log.
CI for AI agents: record real runs, replay them against prompt and model changes, and catch regressions before they ship.
Live Model Context Protocol server that makes this failure registry queryable by any agent.
Streaming investigation assistant that tool-calls the MCP server and answers with cited real case IDs.
Self-hostable prompt-injection challenge: ten levels stack real defenses and a server-side judge scores every crack.
Tooling & evals
Security scanner for MCP servers, like npm audit for the Model Context Protocol: 18 rules, SARIF output.
Security scanner for agent skills: flags prompt-injection, dangerous shell, secret access, and exfiltration before you install a Claude/agent Skill. 31 rules, SARIF.
Trims AI-context bloat: finds the files ballooning your Claude Code, Cursor, and Codex token cost and writes the ignore files to cut it.
Prompt and agent regression CI: the build fails when your prompt gets dumber, with a per-PR quality delta.
Context-window profiler for agents: shows what is eating your context per turn and recommends what to cut.
Verifiable, tamper-evident receipts for RAG answers: Merkle inclusion proofs and Ed25519 signatures.
Observability for AI agents: parses session transcripts with no instrumentation and flags results you should not trust.
Evaluation for voice agents: catches misheard numbers, missing confirmations, and blown latency budgets.
Permission-aware RAG where the access-control check lives inside the retrieval query; CI fails on any leak.
Multi-tenant hybrid search on Qdrant: payload tenant isolation, dense and sparse RRF, HNSW tuning, and benchmarks.
Built by Royal Simpson Pinto. Open source on GitHub.