Skip to content

Back to the catalog

prempti
github.comrepo

prempti

Prempti is an experimental tool from the Falco project that intercepts tool calls made by AI coding agents such as Claude Code and Codex, evaluating shell commands, file reads/writes, web fetches, and MCP calls against customizable Falco YAML rules before execution. It runs in guardrails mode (deny/ask/allow enforcement with LLM-friendly denial explanations) or monitor mode (log-only), ships a default ruleset covering credential theft, sandbox-disable attempts, exfiltration, persistence, and MCP/skill poisoning, and provides a premptictl CLI plus a rule-authoring Claude Code skill. It explicitly notes it is not a sandbox or OS-level containment and is meant to complement, not replace, least-privilege environments.

TL;DR Falco-powered guardrails and audit trail that intercept and police tool calls made by AI coding agents like Claude Code and Codex.
Document

What it is

Prempti (Falco Security, "experimental preview") is a cooperative policy-and-visibility layer for AI coding agents. The existing summary covers its features; the page's actionable content is its explicit threat model, its install/migration constraints, and platform-specific gotchas.

Trust model — the reason to read before acting

The page is emphatic that this is not a security boundary. Guardrails work because the agent cooperates: on a deny or ask verdict the agent receives an "LLM-friendly explanation" and adapts. Prempti explicitly states it "does not contain a determined adversarial agent" and is "not a sandbox, OS-level security, or a substitute for least-privilege environments or system hardening." It is meant to run alongside sandboxing and containment, not replace them. If your threat model includes a compromised or adversarial agent, this tool does not address it.

Preconditions / migration constraint

  • If you already run coding-agents-kit: Prempti does not migrate or remove it. Uninstall coding-agents-kit first, or you get duplicate services and stale Claude Code hooks.
  • Cross-platform: Linux, macOS, Windows on x86_64 and aarch64.

What to do

Install per platform. macOS (universal .pkg), scripted/non-interactive:

installer -pkg prempti-<version>-darwin-universal.pkg \
          -target CurrentUserHomeDirectory

Binaries are not code-signed — Gatekeeper may block them on first run. Allow via System Settings > Privacy & Security, or clear quarantine across the whole tree (both bin/ executables and the share/ plugin library can be flagged):

xattr -dr com.apple.quarantine ~/.prempti

Linux (installs to ~/.prempti/, starts a systemd user service, registers the hook):

tar xzf prempti-<version>-linux-x86_64.tar.gz
cd prempti-<version>-linux-x86_64
bash install.sh

Windows: run the MSI matching your CPU — prempti-<version>-windows-x64.msi (Intel/AMD64) or prempti-<version>-windows-arm64.msi (ARM64). The x64 MSI runs under emulation on ARM64 but the native ARM64 build is preferred.

Verify:

~/.prempti/bin/premptictl status
~/.prempti/bin/premptictl hook status
~/.prempti/bin/premptictl health

Switch enforcement level at any time with premptictl mode <guardrails|monitor> — monitor evaluates and logs rules without acting on the agent, useful for rule tuning before enforcing.

Caveats

Early preview: interfaces and behavior may change between releases. Unsigned binaries mean OS integrity/notarization checks will flag it. As above, it is a cooperative layer, not containment.

Organisations named here

  • falcosecurity/prempti

Technologies named here

  • Claude Code
  • Prempti
  • Falco
  • Codex
  • premptictl
  • Linux
  • coding-agents-kit
  • macOS
  • prempti-falco-rules
  • Windows
  • Model Context Protocol

Techniques named here

  • Real-time tool-call interception
  • Guardrails mode enforcement
  • Monitor mode observation
  • Pipe to shell interpreter
  • Compiled binary evasion of command-level monitoring