Agent Safety & MCP Security
AI agents now run shell commands, call MCP tools and push code. These listings help you decide what an agent may do, check MCP servers before you install them, contain where agents run and what they spend, and keep an auditable record afterwards. 10 tools listed.
withHuman
Open-source human approval layer that intercepts AI agent tool calls, applies rules, and routes risky actions to reviewers.
MELRA
Open-source local MCP runtime that checks AI-agent tool calls against policy, gates risky actions on approval, verifies results and parks uncertain outcomes instead of blindly retrying after timeouts. Apache-2.0.
The MCP Census
Checks public evidence for an MCP server and returns PASS, REVIEW, BLOCK or UNKNOWN before you install or call it; lookup, preflight, whole .mcp.json config audits, a change feed, REST API and its own MCP endpoint, no account needed for stats.
CybeDefend
Security and business rules for AI coding agents: one command installs VibeDefend into Claude Code, Cursor, Copilot or Codex to scan each diff and block risky actions, with SAST, SCA, secrets and IaC scanning; free plan.
SIwithin
Runs Claude Code or OpenAI Codex CLI unattended on a brief, fenced to the folders you give it and under a per-run dollar budget enforced by a loopback API proxy, leaving one ledger line per run. Single Python file, standard library only, macOS only, Apache-2.0.
Lights-on Software Factory
Open-source (MIT) boilerplate for running Claude Code as a software factory: a sandboxed container, a deterministic bash-command regulator, lint hooks for custom rules and a multi-agent workflow generator.
Buddy Sandboxes
Isolated Linux VMs for AI coding agents from Buddy: passwordless sudo, Docker, SSH, a 50 GB disk that survives restarts and public HTTPS endpoints, driven from a CLI, SDK, REST API or MCP tool call.
CoreSpeed
One MCP endpoint that gives Claude Code, Codex or Cursor your apps, your team's shared memory and built-in tools, with spend caps, approvals and an activity log; free plan, Pro $20/month.
Tenbri
Open-source Agent Decision Record: a signed, append-only, hash-chained record of every decision an AI agent makes, verifiable offline; one line instruments Claude, OpenAI and LangGraph agents, with Python and TypeScript SDKs and plain files on storage you control. Apache-2.0, written in Rust.
iFixAi
Independent auditing for AI agents: finds where an agent hides information, overrides decisions or takes unauthorised actions behind a successful result, with recurring audits and reports per agent.