#agent-safety

Control Claude Code in VS Code from Telegram: read sessions, answer questions, approve or deny commands, and keep coding agents moving from your phone.

Self-host secrets manager with an AI-safety layer. Aliases instead of values; AI agents never see real credentials. Cloud option coming.

Border Guard — AI-native territory sovereignty & self-evolving security OS. 6-module D-S fusion engine, trajectory detection, anchor detection, fission engine, territory adjudicator. Full design + Python implementation.

Evidence router & policy engine for coding agents — enforces source choice and proof-of-use through Claude Code hooks, including routes to MCP tools. Local-first; no LLM in the hook loop.

Hard spend caps and a kill-switch for the non-LLM actions your AI coding agent takes. A drop-in Claude Code hook.

Behavioral fuzz-testing engine for AI agents — policy violations, tool-call verification, regression tests. Adapter #1: OpenAI Agents SDK.

Git time-travel for AI coding agents. Checkpoint and rewind Claude Code, Codex, aider, or any command—without touching HEAD, staging, or your branch.

Find the safety rules your AI agent prompt forgot. Offline linter — 23 controls, CLI + Action + browser. No API key.

How far does this command reach? Guards AI coding agents against shell commands that leave your machine. Zero dependencies.

Scanner for agent tool output: flags text that mimics harness messages, overrides instructions, spoofs authority, or hides directives in HTML. Claude Code skill + CLI.

Spring AI starter for Cycles — runtime budget and action authority for Spring AI agents

Deterministic guard hooks for Claude Code and Codex CLI: ten hooks, 1,785 paired must-fire/must-not-fire tests, zero dependencies, one-command install. Refuses 'done' without proof, piped-away git verdicts, off-scope edits, secret writes and config tampering.

Machine-readable preflight for software agents: pass the next unsent tool call; receive ≤2 revision-pinned failure cards or typed UNKNOWN. Offline/zero-install; never authorizes execution.

Two Claude skills for agents that drive web panels. browser-account-guard proves the right account AND the right target before acting; flaky-browser-ui handles tabs that die, clicks that never register, and knowing when to abandon the UI for an API.

Portable, deterministic controls for agent actions. Experimental open-source specification and reference implementation.

Trust-bounded acceptance harness for Claude coding agents — sanitized workspaces, typed tools, isolated oracle evaluation, evidence receipts.

GitHub Action that lints CLAUDE.md-style agent constitution files for missing operational guardrails — spend limits, injection defense, escalation paths, secrets rules