Skip to content

Latest commit

 

History

852 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

VETTO — Daemon-less kernel sandbox for AI coding agents

Run Claude Code, Codex, and Cursor unattended with zero credential-leak anxiety.
Enforced directly by Linux Landlock & Seccomp. No Docker, no root, no daemon.

Release npm version CI Platforms License Fail-Closed
cargo-deny cargo-audit gitleaks


Vetto Demo


Stop Fearing --dangerously-skip-permissions

AI coding agents write impressive code, but running them unprompted on your local workstation is terrifying: a single hallucination, rogue bash loop, or prompt injection can exfiltrate your ~/.ssh keys, wipe your home directory (rm -rf ~), or leak .env secrets.

VETTO wraps Claude Code, Codex, Antigravity, Cursor, and Aider in an OS-level kernel sandbox before the agent process starts:

  • Zero Credential Theft: ~/.ssh, ~/.aws, ~/.gnupg, and .env* are physically unreadable by the agent.
  • Zero Destructive Writes: Agent file modifications are strictly confined to your project root and /tmp.
  • Zero Performance Penalty: ~4ms startup overhead, 0 MB idle RAM, unprivileged execution without Docker.

Installation

Choose your preferred installation method:

# Standalone POSIX script (Linux, macOS, WSL2 — zero dependencies)
curl -fsSL https://raw.githubusercontent.com/shleder/vetto/main/install.sh | bash
# Via npm (global binary for Node.js environments)
npm install --global @shledery/vetto
# Via Cargo (crates.io — compiled from source)
cargo install vetto --locked
# Via Homebrew (macOS & Linux)
brew tap shleder/vetto
brew install vetto
# Or directly: brew install shleder/vetto/vetto
# Via Arch Linux AUR (Arch, Manjaro, EndeavourOS)
yay -S vetto          # Or: yay -S vetto-git
# Or install built package directly with pacman:
sudo pacman -U vetto-0.2.23-1-x86_64.pkg.tar.zst
# Via Chocolatey (Windows)
choco install vetto

# Via Scoop (Windows)
scoop install https://github.com/shleder/vetto/releases/download/v0.2.23/vetto.json
# Run one-off without installing (npx)
npx @shledery/vetto doctor
# In GitHub Actions CI (unattended agent runs, evals, SWE-bench)
- uses: shleder/vetto@v0.2.23

Prebuilt standalone archives with SHA256 checksums and CycloneDX SBOMs for all architectures (x86_64, aarch64, Windows .zip, Linux/macOS .tar.gz) are published on GitHub Releases.

Every release archive is signed with minisign (key id 75ECEC9B5080C590, public key: packaging/release.pub). Verify before running:

# one-time: install minisign (cargo install minisign / apt install minisign / brew install minisign)
minisign -V -p packaging/release.pub -m vetto-linux-x86_64.tar.gz
# → "Signature and comment signature verified"

Every release binary is additionally attested with SLSA Build Provenance (Level 3) via GitHub Sigstore. Verify provenance (guide):

gh attestation verify vetto-linux-x86_64.tar.gz --owner shleder
# → "Verification succeeded!"

Quick Start

Protect your workstation and run any AI coding agent unattended in seconds:

1. Enable Your Agent

# Wrap any agent of choice (creates transparent zero-overhead shims):
vetto enable opencode      # OpenCode CLI
vetto enable claude        # Claude Code
vetto enable codex         # OpenAI Codex CLI
vetto enable windsurf      # Codeium Windsurf
vetto enable goose         # Block Goose
vetto enable cursor        # Cursor Agent
vetto enable aider         # Aider

Creates transparent, zero-latency shims in ~/.vetto/shims/ and configures shell PATH priority.

Supported AI Coding Agents (20 Native Presets)

Vetto provides tailored sandbox presets, project marker auto-detection, and zero-config network allowlists for 20 AI coding tools:

  • opencode (OpenCode) · claude (Claude Code) · codex (OpenAI Codex) · gemini (Google Gemini)
  • antigravity (Google Antigravity / agy) · cursor (Cursor Agent) · aider (Aider) · cline (Cline)
  • copilot (GitHub Copilot CLI) · windsurf (Codeium Windsurf) · continue (Continue CLI) · goose (Block Goose)
  • openhands (All-Hands OpenHands) · swe-agent (SWE-agent) · plandex (Plandex) · mentat (Mentat)
  • gpt-engineer (GPT Engineer) · devin (Cognition Devin) · crust (Crust AI) · amp (Amp AI)

2. Run Completely Unattended

Launch your agent with full autonomy:

# OpenAI Codex
codex exec --full-auto

# Claude Code
claude --dangerously-skip-permissions

# Antigravity CLI
antigravity run --autonomous

# Aider
aider --yes

# Or run any agent directly inside vetto:
vetto -- claude
vetto -- aider
vetto -- cursor
vetto -- <agent> [args...]

Files outside the workspace are blocked, host credentials (~/.ssh, ~/.aws, .env) are masked, and network egress is locked down to provider APIs.

To check wrapped status or unwrap at any time:

vetto enable --status      # Check all active shims
vetto disable <agent>      # Unwrap agent (e.g. vetto disable codex)

Why Not Docker?

Containers were designed for packaging backend microservices—not for interactive developer coding loops. Vetto enforces OS-level kernel confinement directly around your host processes:

Dimension VETTO Docker Containers Why It Matters
Startup Overhead ~4ms (CI-gated spawn overhead) 3.5s – 8.0s Subagents and test loops execute with zero perceptible latency
Daemon None (zero background processes) dockerd background service No background daemon to crash, stall, or consume idle resources
RAM Overhead 0 MB 1.5 GB+ (VM / daemon engine) Leaves all workstation RAM free for compilation and local models
Permissions Unprivileged (no root / no sudo) Root / docker group (root-equivalent) Completely eliminates root-escalation attack surface on your host
Host File Sync Native Filesystem (instant) Volume mounts (slow I/O, UID sync bugs) Edits, hot-reloading, and git diffs reflect immediately
Kernel Barrier Linux Landlock + Seccomp-BPF Namespaces + cgroups In-process confinement applied before execve, strictly fail-closed
Network Egress Per-Domain Allowlist (api.anthropic.com) All-or-nothing bridge Prevents unauthorized data exfiltration without breaking inference

Platform Support

Security tooling frequently makes misleading cross-platform claims. Vetto rejects snake oil and is architecturally transparent about what each operating system kernel can and cannot enforce unprivileged.

Vetto establishes an immutable 3-tier boundary architecture:

  • Tier 1 (Production-Grade): Linux (Native) and Linux (WSL2). Full kernel isolation via Landlock LSM (ABI v1–v6), Seccomp-BPF (UnixOnly / AgentMin), rlimit ceilings (AS, NPROC, CPU, FSIZE), sticky NO_NEW_PRIVS, dedicated process groups, private Mount/PID/Network namespaces, and full /proc nonce tree-sweep.
  • Tier 1 via VM (Uniform path): macOS → Linux VM (mac-vm, Virtualization.framework) and Windows → WSL2 (guest vetto + sync-back). Same Tier-1 stack inside the guest; default on macOS/Windows when the VM is configured, fail-closed with an actionable message when it is not.
  • Tier 2 (Experimental, maximum-achievable): macOS native (Seatbelt). Seatbelt write isolation, network lockdown (--net off), pgroup host verification, and group-death sweep. Reads stay broad — the SBPL matrix proves only Shape A runs (all fragmented shapes SIGABRT incl. static Go), so narrow read-deny is closed until Apple fixes dyld (#62). For full read isolation, run inside OrbStack or a Linux VM.
  • Tier 3 (Preview, maximum-achievable): Windows native (AppContainer/LPAC + Job Objects). Job Objects tree containment with kill-on-close (JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE), AppContainer process and network-off (--net off) boundary, policy-conditional ceilings. No seccomp or BPF on Windows; syscall filtering and exec-root isolation remain unsupported. Use WSL2 for full Tier 1 kernel boundary.

Canonical 5-Factor Capability Matrix

Platform / Tier Filesystem Write Filesystem Read Network Namespace Process Reaping Secret Overlays per OS Assurance Status
Linux (Native)
Tier 1 (Production)
100% Kernel Deny (Landlock ABI v1–v6 + R/O Mounts) 100% Scoped Read (Landlock VFS Inode checks, ~/.ssh / .env blocked) Yes (CLONE_NEWNET, loopback-only + local TCP/TLS broker) 100% PID Namespace (CLONE_NEWPID init teardown + PR_SET_PDEATHSIG + full /proc nonce tree sweep) Yes (tmpfs mode-000 and /dev/null bind-mounts over secrets) Tier 1 (Proven): Complete hardware & kernel isolation (Landlock, seccomp UnixOnly/AgentMin, rlimits, sticky NO_NEW_PRIVS)
Linux (WSL2)
Tier 1 (Production)
100% Kernel Deny (Landlock via WSL2 Linux Kernel) 100% Scoped Read (Landlock VFS Inode checks) Yes (CLONE_NEWNET inside WSL2 VM) 100% PID Namespace teardown + /proc sweep Yes (tmpfs mount overlays inside WSL2) Tier 1 (Production): Recommended path for Windows workstations
macOS (Darwin)
Tier 2 (Experimental)
100% Locked (Seatbelt SBPL (allow file-write*) to workspace & /tmp) Broad Reads (System / read due to dyld bug; tail deny on known secrets) No (Unsupported by Darwin; --net=off via SBPL (deny network*)) pgroup host sweep (Watchdog kqueue pdeath_watch + group-death SIGKILL sweep) No (VFS overlays unavailable unprivileged; SBPL static deny only) Tier 2 (Experimental): Write confinement and --net=off network lockdown. Unprivileged read-isolation and syscall filtering unsupported (use OrbStack/Linux VM for full read isolation)
Windows Native
Tier 3 (Preview)
Workspace Only (AppContainer DACL + LPAC S-1-15-2-2 write grants) ACL Fallback (AppContainer default-deny; partial token restriction) No (Network namespaces unavailable; --net=off via AppContainer caps) 100% Job Object (JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE terminates process tree) No (No unprivileged mount namespaces; fails closed on collision) Tier 3 (Preview): Process guardrails with kill-on-close and --net off; no seccomp/BPF or exec-root isolation. Use WSL2 for full kernel boundary
Windows Sandbox
Tier 3 (VM Isolated)
VM Isolated (Dedicated virtual disk, mapped read-write folders only) VM Isolated (Host secrets never mapped into .wsb specification) Virtual Switch (Hyper-V vSwitch disabled under --net=off) VM Teardown (Hyper-V VM instance termination) Full Isolation (Physically separated filesystem in disposable VM) Disposable VM: Hardware-virtualized container (requires Hyper-V)

Honest Status

Fail-Closed Principle

Vetto strictly enforces a fail-closed contract across all platforms. If a requested security boundary cannot be guaranteed by the current operating system kernel or runtime environment, Vetto exits immediately with code 125 (EXIT_FAIL_CLOSED). Vetto never silently degrades to an unconfined or insecure execution state.

The Honest macOS Disclosure

Apple has deprecated SBPL (sandbox-exec) and Darwin kernels impose severe constraints on unprivileged file-read denial:

  • Modern versions of Apple's dynamic linker (dyld) on macOS 13, 14, and 15 crash with SIGABRT when SBPL file-read rules are fragmented across multiple discrete path clauses.
  • Vetto transparently tracks this platform defect via vetto doctor under the sbpl-read-fragment probe.
  • Rather than crashing agent workflows or manufacturing illusory read security, Vetto on macOS grants broad system reads while strictly enforcing 100% filesystem write lockouts, --net=off network isolation, and process supervision (pgroup host verification and group-death sweep).
  • Recommendation: If your threat model requires 100% hardware-enforced kernel read-denial of host credentials (~/.ssh, ~/.aws, .env) on macOS, execute Vetto inside OrbStack, a lightweight Linux VM, or Docker devcontainers where Linux Landlock and mount namespaces are available.

The Honest Windows Disclosure

Native Windows isolation uses Job Objects and Less Privileged AppContainers (LPAC). While Job Objects guarantee 100% child process termination on close (JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE) and AppContainer isolates network egress (--net off):

  • Windows kernels do not support unprivileged mount namespaces, seccomp, or BPF. Syscall filtering and exec-root isolation remain unsupported.
  • Native Windows is designated Tier 3 (Preview).
  • For production-grade Tier 1 protection on Windows workstations, use WSL2 (wsl -- vetto ...), which provides the native Linux kernel Landlock LSM, seccomp-bpf, and namespace isolation stack.

Scope Closure: Issues #26, #62, #63

Issue #26 tracks the honest per-OS matrix. #62 (macOS read-isolation) is closed as maximum-achievable: the SBPL matrix proves Shape A is the only runnable shape. #63 (Windows hardening) is closed as maximum-achievable: AppContainer/LPAC + Job at ceiling, WFP admin-gated, signing status explicit. Vetto permanently repudiates ungrounded claims of cross-platform parity:

  1. Platform capabilities are strictly tiered (Tier 1 Linux, Tier 2 macOS, Tier 3 Windows).
  2. All capability claims are continuously verified in CI via automated red-team matrices and diagnostic doctor probes.
  3. Pull requests or features claiming parity without underlying OS kernel enforcement proofs will be rejected.

Ecosystem Integration Guides

Vetto natively integrates with modern AI coding workflows:

Agent / Tool Guide & Details Setup Command
Claude Code Claude Code Guide · Unprompted mode, PreToolUse hook, Anthropic API allowlist vetto enable claude
Cursor Cursor Guide · Agent & Composer sandboxing, terminal execution, storage masking vetto enable cursor
Cline Cline Guide · VS Code extension terminal task isolation, zero-config shims vetto hook install
Aider Aider Guide · Zero-config network allowlists, git protection, automated tests vetto enable aider
OpenCode & Codex OpenCode Guide · CLI runners, subagent supervision, and model sandboxing vetto enable codex
Claude Desktop & Codex Desktop Desktop Integration Guide · Native MCP server (vetto mcp), terminal shims, sandboxed subtools vetto mcp · vetto enable

Boundary Verification & Audit

Trust nothing—verify the sandbox boundary and audit past operations:

# Probe running kernel capabilities (Landlock ABI, seccomp, userns)
vetto doctor

# Show concrete OS remediation steps for missing primitives (Seatbelt, Job Objects, Hyper-V)
vetto doctor --fix

# Additionally verify that every denied path is truly unreachable in a throwaway sandbox
vetto doctor --probe

# Run active leak battery (tests secret paths and loopback isolation)
vetto verify

# Inspect effective policy and test path blocks
vetto policy explain
vetto policy explain --why ~/.ssh/id_rsa

# Inspect intercepted security violations and blocked paths from past sessions
vetto audit
vetto audit --latest

# End-of-session security recap (top denied paths, egress split, intent)
# printed automatically after every session; post-hoc on demand:
vetto audit --latest --recap

Dynamic Policy Grants (No Manual TOML Editing)

When an agent is blocked from accessing a legitimate project path, Vetto prints an immediate grant command:

vetto allow ./vendor                    # Grant read+write to a folder
vetto allow --read-only /usr/share/doc  # Grant read-only access
vetto allow --net registry.npmjs.org    # Allow egress to a package registry
vetto deny ~/.aws/credentials           # Explicitly mask a secret file

Advanced CLI Execution

Beyond transparent vetto enable shims, you can run one-off commands or custom agents directly:

# Auto-detect agent in current workspace and run inside sandbox
vetto

# Explicit command supervision
vetto -- claude -p "fix failing tests"
vetto -- aider --model sonnet
vetto -- codex exec "refactor auth module"

# Security presets: balanced (default) | paranoid | yolo
vetto --preset paranoid -- npm test

# Network modes: off (default) | ask | allowlist:<domains> | strict:<host:port>
vetto --net allowlist:api.anthropic.com,github.com -- cargo check
vetto --net strict:github.com:22 --git-ssh -- git fetch origin

# Explicit sandbox backend: auto (default) | process | win-sandbox (Hyper-V disposable VM)
vetto --backend auto -- npm test

# Terminal UI modes: statusline (default) | full (dashboard) | none (headless)
vetto --tui full -- claude

# Non-interactive CI mode (implies --tui=none and emits structured JSON summary)
vetto --ci -- npm test

# Shadow mode: policy layer logs "would deny" during preflight without blocking
# (Note: kernel sandboxes Landlock/Seatbelt cannot be shadowed)
vetto --shadow -- cargo test

# Output detailed HTML / SARIF audit reports
vetto --report html,sarif --jsonl session.jsonl -- cargo test

Note on --tier: Vetto automatically detects and negotiates the highest available sandbox tier based on kernel capabilities (Tier 1 Linux, Tier 2 macOS, Tier 3 Windows). --tier is not a top-level runner flag (use --backend for backend selection); --tier exists on vetto shell-env --tier for shell exports. On Linux, test degraded tiers via VETTO_FORCE_TIER.


Model Context Protocol (MCP) & Multi-Agent

Sandboxing MCP Servers (vetto mcp wrap)

Wrap and isolate third-party Model Context Protocol (MCP) servers (e.g. filesystem tools, API connectors):

# Wrap an external MCP server binary:
vetto mcp wrap --allow ./data --allow-read /usr/share --net off -- <server-binary> [args...]
  • Windows Path Hardening: Automatically resolves and grants read access to canonical %SystemRoot% (C:\Windows), System32 (verified against cmd.exe/kernel32.dll to prevent ENV-POISON attacks), %ProgramFiles%, and temporary directories.
  • Platform Network Boundary: Network egress relay requires Linux network namespaces (CLONE_NEWNET). On non-Linux platforms (macOS and Windows), vetto mcp wrap requires --net off (default); other network modes fail closed.

Multi-Agent Orchestration (vetto multi)

Run multiple AI coding agents concurrently, each in an independent sandbox with an interactive split-pane dashboard:

# Run multiple agents defined in a TOML manifest:
vetto multi --manifest vetto-agents.toml

# Or specify multiple named agents directly:
vetto multi --agent coder=claude --agent reviewer=codex
  • Platform Constraint: The multi-agent orchestrator is fully supported on Unix platforms (Linux and macOS). On Windows, vetto multi is blocked fail-closed with UnsupportedPlatform: multi-agent.

What Vetto Deliberately Excludes

  • No background daemon — zero idle CPU, zero RAM consumption, no service to stall or crash.
  • No root / sudo — runs completely unprivileged; cannot escalate host permissions.
  • No TLS interception — zero MITM, no custom root certificate authority; moves opaque bytes only.
  • No telemetry or tracking — completely private by default. No telemetry or project/user data is ever transmitted. The only network calls vetto itself initiates are short, non-blocking version checks against the npm registry and GitHub Releases (24h cache, 2s timeout, silent offline via cache). Self-update never runs unless explicitly opted in, and never in CI (VETTO_NO_SELF_UPDATE=1 disables everything update-related).
  • No Docker dependency — instant ~4ms startup directly on your native OS kernel.

Deep Architecture & Kernel Enforcement (Click to expand)

1. The Fail-Closed Discipline

Vetto puts the agent process inside an OS-level sandbox before the agent process starts. The foundational guarantee of Vetto is fail-closed execution:

If the requested boundary cannot be established on the current host, vetto exits immediately instead of starting the agent. There is no fallback to an unconfined process anywhere in the codebase.

2. Linux Landlock LSM & Seccomp-BPF

  • Landlock ABI Negotiation: Automatically negotiates Landlock ABI versions 1 through 6 with the running kernel. Landlock restricts filesystem operations (open, read, write, unlink, rename) directly in kernel space.
  • Seccomp-BPF: Enforces fine-grained syscall restrictions before execve. Disallowed syscalls receive EPERM or ENOSYS.
  • Secret Masking (display_only_deny): Because Landlock is a pure allowlist and cannot subtract subpaths from an allowed directory tree, Vetto masks secret files (such as ~/.ssh, ~/.aws, ~/.gnupg, .env, tokens) by mounting an empty tmpfs or /dev/null over them on the Linux full tier, or by carving them out of generated read allowlists.
  • Path Resolution: Symlinks and globs are expanded to concrete filesystem paths before rules reach the kernel. Patterns never reach the kernel.

3. In-Process Network Relay Broker

  • Network Namespaces: When network filtering is enabled (--net=allowlist:...), Vetto isolates the child process in a dedicated Linux network namespace with only a loopback device.
  • Broker Relay: Outbound TCP connections route through an in-process local CONNECT/SOCKS broker.
  • DNS Validation & Anti-Rebinding: The host broker performs DNS resolution itself and pins addresses per rule. Any DNS response resolving to private IP ranges (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 169.254.169.254) is rejected.
  • Zero TLS MITM: The broker moves opaque bytes. There is no TLS interception, no custom certificate authority, and no MITM proxy.

4. Recursion Barriers & Transparent Shims

  • When vetto enable <agent> creates shims in ~/.vetto/shims, recursion barriers (VETTO_WRAPPED, VETTO_SANDBOXED, VETTO_SHIM_ACTIVE) guarantee that subagent tool invocations (such as an agent invoking git, python, or nested compiler toolchains) resolve directly to real host binaries without recursive supervisor overhead or infinite loops.
  • vetto enable refuses to overwrite non-Vetto binaries without --force.

5. Policy Layer Hierarchy

Policies merge in a deterministic, strict hierarchy where every TOML struct rejects unknown fields:

Host Global (/etc/vetto/config.toml)
  └── User Global (~/.vetto/config.toml)
        └── Built-in Profile (default, strict, paranoid)
              └── Agent Preset (claude, cursor, aider, cline, codex)
                    └── Project Policy (./vetto.toml + policy.d/)
                          └── Local Override (./vetto.local.toml)
                                └── CLI Overrides (--allow, --net, --limits)

Session Rescue & Diagnostics (Click to expand)

Recover interrupted, frozen, or corrupted agent sessions without losing progress:

# Scan recent sessions
vetto rescue --json scan --limit 25

# Diagnose Claude Code or Cursor sessions
vetto rescue --adapter claude diagnose <session-id>
vetto rescue --adapter cursor snapshot <session-id> --output ./recovered.jsonl

# Rollback a failed repair
vetto rescue rollback --receipt <receipt-path>

Adapters supported: claude, cursor, codex. Snapshots are verified with SHA-256 and created strictly outside the original state root.


Documentation


License

Apache-2.0 — see LICENSE and THIRD_PARTY_NOTICES.md.

About

Daemon-less OS security boundary & sandbox for AI coding agents (Codex, Claude Code, Cursor, Aider)

Topics

Resources

Security policy

Stars

21 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages