Meerkat
Local-first, agent-agnostic CLI security wrapper: auto-approves safe commands, blocks risky ones, logs every decision, so AI coding agents run without blind trust.
Vibe coding without blind trust.
Autonomous coding agents, Claude Code, Codex, aider, goose, execute real terminal commands on your machine. Letting them run unsupervised is dangerous; approving every command by hand defeats the point of automation.
Meerkat sits between the agent and the shell: a secure-by-default local policy runner that keeps the machine awake, auto-approves what is safe and scoped, blocks what is not, and records every decision.
Agent-agnostic policy at the hook layer.
It wraps any command-line tool, `meerkat run -- npm test`, or installs directly into Claude Code as a `/meerkat` slash command with PreToolUse / SessionStart / Stop hooks. Every session auto-approves safe shell commands per a `meerkat.yml` policy and blocks `sudo`, `curl`, `git push origin main`, and `--force` pushes.
It hardens with OS-level backends, Seatbelt on macOS, bubblewrap on Linux, WSL2 on Windows, when opted in, while staying honest about what it is: a policy runner, not a kernel sandbox.
Be explicit about the threat model.
The README documents exactly what is and is not covered, supervised Claude Code sessions and direct wrappers are; subprocess behavior inside already-approved commands, kernel exploits, and packet-level egress are not. For high-risk workloads it points users to containers or VMs rather than overpromising.
Decisions are logged as JSONL so any auto-approval or block is auditable after the fact.
Installable in one line, shipping with CI/release automation.
Distributed as a single Go binary via an install script, `npx meerkat-cli`, or `go install`, with GitHub Actions CI and release pipelines and branch/tag protection, built like a tool meant to be used, not a demo.