belay

<div align="center"> <img src=".github/readme-assets/lockup.gif" width="380" alt="Belay">

Detect · Block · Notify — a defense and monitoring layer for AI coding agents.

</div>

Your coding agent has full shell access. One prompt injection in a GitHub issue, one hallucinated rm -rf, one MCP tool response that isn't what it claims to be — and it's not a bug report anymore, it's your .env in a stranger's webhook or your home directory gone.

Belay is an open-source, local-first security layer that gates every tool call an AI coding agent makes — blocking destructive commands, secret leaks, and dangerous MCP tool calls in under 100ms, with no LLM in the decision path, no cloud round-trip, and no phone-home by default.

It auto-detects the agents you already run, blocks dangerous actions outright, escalates ambiguous-but-risky ones to you for a one-tap Allow / Deny (in the terminal, the desktop app, or your phone), and notifies you when something happens.

<div align="center"> <img src=".github/readme-assets/activity-demo.gif" alt="Belay's live Activity feed showing real attack attempts being blocked in real time — destructive rm -rf, credential exfiltration, fetch-and-exec droppers, all denied" width="720"> </div>

---

What it stops

• Secret exfiltration — .env files, API keys, SSH keys, and credentials belonging to other AI tools on the same machine
• Destructive commands — rm -rf /, curl … | sh, fetch→chmod→exec droppers — including common evasion shapes (wrapper prefixes, compound chaining, heredocs, split/long flags)
• Reverse shells and command-and-control egress
• MCP tool-response injection markers — flags embedded "ignore previous instructions"-style content in a tool's response, correlated with any secret-holding session that follows
• Supply-chain, persistence, privilege-escalation, and recon patterns
• Malicious agent skills — install-time gating plus ongoing drift detection, so a skill that's benign at approval time and swapped later gets re-caught

<div align="center"> <img src=".github/readme-assets/alerts.jpg" alt="Belay's live Alerts feed showing a secret redacted from an MCP response and an injection marker detected in an MCP tool response" width="720"> </div>

Every rule is tagged with the OWASP Top 10 for Agentic Applications (from OWASP's Agentic Security Initiative), OWASP LLM Top 10, and MITRE ATLAS, and the static scanner emits SARIF 2.1.0 straight into your CI's code-scanning tab.

Known limitation, stated plainly: command-gate detection is pattern-based, not a full sandbox — an interpreter-wrapped call that never surfaces a recognizable shell substring (e.g. a destructive action expressed purely through a scripting language's own API) can still slip through, and a bare network call to a cloud metadata endpoint outside a skill's context isn't yet a dedicated command-gate rule. Belay is defense-in-depth, not a guarantee.

Supported agents

Most agent-safety tools protect one agent. Belay auto-detects and correctly instruments 11 — using the right native mechanism per agent, not a generic shim:

| Agent | Interception | |---|---| | Claude Code | native hook | | Codex | native hook | | Cursor | native hook | | Hermes | native hook | | OpenClaw | exec policy | | Gemini CLI | config policy | | Goose | config policy | | Cline | MCP proxy | | Roo | MCP proxy | | Antigravity | MCP proxy | | Opencode | plugin |

More agents land regularly — run belay detect to check yours, and see MCP proxy below for wrapping any MCP server Belay doesn't natively recognize yet.

<div align="center"> <img src=".github/readme-assets/agents.jpg" alt="Belay's Agents view showing claude-code and openclaw detected, with their config paths, MCP servers, and installed skill counts" width="720"> </div>

Approve from your phone

Step away from your laptop — when a call is an Ask, Belay can send the prompt to a chat app and take your Allow/Deny reply back. Two-way on Telegram, Discord, WhatsApp, Matrix, Mattermost, and Slack; notify-only on ntfy, Teams, WeCom, and webhooks. Enrolment is a one-time pairing code, and it's default-deny — only enrolled approvers can approve, and a prompt that goes stale is auto-denied, never auto-allowed.

<div align="center"> <img src=".github/readme-assets/telegram-approval.gif" alt="A phone showing a Belay approval request in Telegram — a CRITICAL fetch-and-exec dropper is denied with one tap, then a MEDIUM RISK cross-agent config read arrives with a plain-English explanation of why it matters" width="320"> </div>

How it works

detect  →  protect  →  serve / desktop app
  │           │              │
  │           │              └─ live status, triage, and Allow/Deny in the desktop UI
  │           └─ gate every tool call: Deny > Ask > Allow (fail-closed)
  └─ find the AI agents installed on this machine

Belay gates each tool call deterministically. A Deny can never be downgraded by the dev-toolchain allowlist; an Ask waits for a human decision and denies on timeout. Run with --observe first to tune in log-only mode before enforcing.

How Belay is different

• Broader than a command blocklist. Tools that only block destructive shell commands don't cover secret exfiltration, MCP tool-call abuse, malicious agent skills, or host-level egress. Belay gates all of them from one rule catalog.
• Gates the tool call, not just the shell. Belay sits at the agent's tool-call boundary across native hooks (Claude Code, Codex, Cursor, Hermes), config policy (Gemini, Goose), and an MCP proxy (any MCP server) — not one integration surface.
• Local-first, not a SaaS. No cloud dependency, no LLM in the decision path, no phone-home by default, self-hostable, and the AGPL-3.0 source is yours to read end to end — inspect exactly what it does before you trust it with shell access.

Features

• Runtime gating at the tool-call boundary — native hooks for Claude Code, Codex, Cursor, and Hermes; config-policy for Gemini and Goose; exec-policy for OpenClaw. PostToolUse redacts secret-shaped strings from tool output.
• MCP proxy (mcp-proxy) — wrap any MCP server so every tools/call is gated; fail-closed. Also scans tool responses for embedded injection markers and alerts, correlated against any secret-holding session.
• Deterministic rule catalog — secrets, egress, destructive, RCE, supply-chain, persistence, priv-esc, recon, config-tamper, plus arm→sink and "lethal-trifecta" session correlation.
• Skill security scanning — a dedicated scanner for AI-agent skills: prompt-injection, SSRF/cloud-metadata theft, credential snooping, tool-description poisoning, and anti-refusal manipulation, checked at install time and on an ongoing directory watch with content-hash-keyed trust so a skill can't be swapped for something malicious after approval without re-triggering review.
• Human-in-the-loop approvals — Allow / Deny in the terminal, the desktop app, or a chat app (see Approve from your phone); timeout denies.
• Explain & Advise — every verdict carries a plain-English "why is this risky / what to do" explanation. An optional AI explainer (OFF by default, bring-your-own-key: local Ollama or a cloud provider) can add a second opinion. It is advisory only — it never makes or changes a decision, and secrets/paths are redacted before anything is sent.
• Static pre-install scanner (scan) — patterns, AST, taint, YARA, and OSV analyzers with provenance-weighted scoring and SARIF output for CI. An optional --llm cascade can filter false positives; with no API keys it runs fully local and heuristic-only.
• Tamper-evident audit — a hash-chained audit log plus evidence build / evidence verify (SHA-256 manifest over findings + SARIF).
• Honeypot canaries — decoy credential files that trip a Critical verdict on read or egress.
• Host control — a native Rust firewall and a bundled per-ecosystem vulnerability DB, with no external tools and no NVD key required. The vuln DB surfaces CISA KEV (known-exploited) badges and EPSS exploit-probability scores, and outbound destinations can be annotated with reverse-DNS + ASN / owner / country (display-only — never gates).
• Desktop app (Tauri 2, Windows/macOS/Linux) — system-tray status, a live dashboard, real-time Alerts feed, per-agent protection detail, and privacy-safe notifications (category only — never your paths or commands).

Install

Linux & macOS (CLI):

curl -fsSL https://dl.belay.secblok.io/install.sh | bash

Windows (desktop app, PowerShell 5.1+):

irm https://dl.belay.secblok.io/install.ps1 | iex

Prefer not to pipe a script for a security tool, understandably — read it first, or grab a direct binary and verify its checksum yourself:

# Read before running
curl -fsSL https://dl.belay.secblok.io/install.sh -o install.sh
less install.sh && bash install.sh

# ...or download a checksummed desktop installer directly
# Windows:      belay-setup-x64.exe
# macOS (arm):  belay-aarch64.dmg     macOS (Intel): belay-x64.dmg
# Linux (.deb): belay-amd64.deb       Linux (AppImage): belay-amd64.AppImage
# https://github.com/SECBLOK/belay/releases/latest

Every download is checked against a published SHA256SUMS/.sha256 before installing, refusing on a mismatch. The Windows and macOS installers are currently unsigned while a code-signing/notarization certificate is pending — Windows SmartScreen and macOS Gatekeeper may warn on first run ("More info → Run anyway" / right-click → Open).

Uninstall anytime with belay uninstall (add --purge to also remove ~/.belay).

Quick start

Once installed, drive Belay with the belay binary already on your PATH (or launch the desktop app, which walks you through the same flow):

# 1. See which agents are installed
belay detect

# 2. Protect one in log-only mode first (tune, no enforcement)
belay protect <agent> --observe

# 3. Enforce
belay protect <agent>

# 4. Run the local backend (the desktop app renders the UI)
belay serve

Review activity with belay status (the last audit rows) and belay logs (a longer tail).

belay serve exposes a local API + SSE backend on 127.0.0.1 — it is not a web dashboard. The desktop app is the user interface. Enable boot-start with belay install-service --enable (systemd / launchd).

Build from source

Contributors can build the single static binary from a checkout:

cargo build --release --bin belay

The primary Linux target is a fully-static musl build (--target x86_64-unknown-linux-musl), which needs the musl toolchain — install musl-tools (Debian/Ubuntu: sudo apt install musl-tools) so musl-gcc is available.

Contributions are welcome — see CONTRIBUTING.md. Contributions require a CLA + DCO sign-off.

Platform support

| Platform | CLI | Desktop app | |----------|-----|-------------| | Linux x86_64 | Released — fully-static musl build, runs on any libc | Released — .deb / .AppImage | | macOS (Intel + Apple Silicon) | Released | Released — .dmg, both architectures | | Windows | Released | Released — NSIS installer (unsigned, cert pe