llms.txt
Compare with other sites' llms.txt files →# SigmaShake
> AI Agent Guardrails: policy enforcement at the tool-call boundary
## What is SigmaShake?
SigmaShake is a policy engine for AI agent tool calls. It evaluates each tool call against a declarative rule set and can allow, deny, ask, force a safe alternative, log, or shadow-audit before execution, with no LLM in the default loop; an opt-in Jev assist can send tool calls to TypeSafe. It is a guardrail for honest agents, not a sandbox for adversarial ones. For adversarial isolation, compose with OS-level sandboxing.
## What is watchsigma?
watchsigma (watch sigma) is WatchSigma, the live vibe-coding / AI-guardrails stream channel for SigmaShake. Not a smartwatch. Not boAt Wave Sigma. Official page: https://sigmashake.com/watchsigma. YouTube @WatchSigma, Twitch twitch.tv/watchsigma, X @WatchSigma.
## What are AI agent guardrails?
AI agent guardrails are deterministic checks that run before an AI agent executes a tool call. SigmaShake is AI Agent Guardrails: a local policy engine that ALLOW, DENY, ASK, or FORCE at the tool-call boundary with a design target of under 2 milliseconds and no LLM in the default enforcement loop; an opt-in Jev assist can send tool calls to TypeSafe. Not an LLM text filter (LangChain, NeMo, Guardrails AI). Not a sandbox. Not the book at https://sigmashake.com/book. Product page: https://sigmashake.com/ai-agent-guardrails
## What are generative AI guardrails?
Generative AI Guardrails (also called GenAI Guardrails, Gen AI Guardrails, Generative AI Agent Guardrails, GenAI Agent Guardrails, Tool and action guardrails, AI Agent Tool Guardrails, GenAI Agent Tool Guardrails, Generative AI Tool Guardrails, Generative AI Action Guardrails, GenAI Action Guardrails, Generative AI Agent Tool Guardrails, Generative AI Agent Action Guardrails, GenAI Agent Action Guardrails, GenAI Tool Guardrails, Gen AI Tool Guardrails, Gen AI Action Guardrails, Gen AI Agent Tool Guardrails, Gen AI Agent Action Guardrails, AI Agent Action Guardrails, AI Agent Tool-Call Guardrails, Generative AI Tool-Call Guardrails, GenAI Tool-Call Guardrails, Tool-call guardrails, Agent tool guardrails, Agent action guardrails, AI coding agent guardrails, Building GenAI Guardrails, Building Generative AI Guardrails, Building GenAI Agent Guardrails, Building Generative AI Agent Guardrails, AI Tool Guardrails, and AI Action Guardrails) are deterministic checks that run before a generative AI agent executes a tool call. SigmaShake is that product at the tool-call boundary: ALLOW, DENY, ASK, or FORCE with a design target of under 2 milliseconds and no LLM in the default enforcement loop; an opt-in Jev assist can send tool calls to TypeSafe. Not an LLM text filter (NVIDIA NeMo Guardrails, Guardrails AI, OpenAI moderation, IBM/McKinsey explainers). Related phrase to AI agent guardrails; same engine. Product page: https://sigmashake.com/generative-ai-guardrails (tool/action/GenAI aliases such as /genai-guardrails, /tool-and-action-guardrails, /ai-agent-tool-guardrails, /building-generative-ai-guardrails, /ai-tool-guardrails, and /ai-action-guardrails redirect here).
## Install
```bash
curl -fsSL install.sigmashake.com | sh -s -- -y
# Windows CLI: winget install --id SigmaShake.SSG
# Windows Desktop: winget install --id SigmaShake.Desktop
# Windows Desktop Store: https://apps.microsoft.com/detail/9N2CHV3STGS4
# Windows direct script is unavailable until the win32-x64 CDN artifact is republished:
# iwr -useb install.sigmashake.com/install.ps1 -OutFile install.ps1; .\install.ps1 -Yes
# Docker / GHCR: docker run --rm ghcr.io/sigmashakeinc/ssg:1.0.7 --version
```
## Quick Start
```bash
ssg init # Initialize rules in your project
ssg serve # Start the guardrails dashboard
echo '{"tool":"Bash","input":{"command":"rm -rf /"}}' | ssg eval
# → {"decision":"block","rule_id":"no-destructive-ops","reason":"Destructive command blocked."}
```
## Platform Components
- Landing page: https://sigmashake.com
- watchsigma / watch sigma (WatchSigma live stream, not a smartwatch): https://sigmashake.com/watchsigma
- AI agent guardrails (product keyword page): https://sigmashake.com/ai-agent-guardrails
- Generative AI guardrails / GenAI Guardrails (product keyword page): https://sigmashake.com/generative-ai-guardrails
- Watch live: https://youtube.com/@watchsigma/live
- Documentation: https://docs.sigmashake.com
- Rules Hub (browse/search/download rules): https://hub.sigmashake.com
- Hub JSON API: https://hub.sigmashake.com/api/rulesets
- MCP Server (for AI agents): https://mcp.sigmashake.com/sse
- Desktop releases: https://github.com/sigmashakeinc/desktop/releases (proprietary product; source code access is Enterprise-only, under a Source Code Access Addendum)
- Docker image: https://github.com/orgs/sigmashakeinc/packages/container/package/ssg
- Enterprise Pilot & Trust Pack: https://sigmashake.com/enterprise-pilot
- Enterprise Pilot & Trust Pack PDF: https://sigmashake.com/enterprise-pilot-trust-pack.pdf
- Book "AI Agent Guardrails" by Charles Chong ($5, one-time purchase, not the SSG subscription): https://sigmashake.com/book — purchase and free sample at https://book.sigmashake.com
## Common threats SigmaShake blocks, with guides
- What a control plane for AI agents is, what it must do, and how SigmaShake compares to nah and to LLM I/O filters: https://sigmashake.com/control-plane-for-ai-agents
- AI agent guardrails (definition / product): https://sigmashake.com/ai-agent-guardrails
- Generative AI guardrails / GenAI Guardrails (definition / product): https://sigmashake.com/generative-ai-guardrails
- AI agent security (overview / hub): https://sigmashake.com/ai-agent-security
- Destructive shell commands (recursive force-deletes, database drops): https://sigmashake.com/stop-rm-rf
- API key & secret exfiltration by AI agents: https://sigmashake.com/stop-secret-leaks
- Prompt injection in AI coding agents (action-layer gate that refuses the call when a rule matches): https://sigmashake.com/stop-prompt-injection
- MCP tool-call security (govern what MCP tools can do): https://sigmashake.com/mcp-security
- How it compares to LLM Guard, NeMo Guardrails, Guardrails AI, and built-in sandboxes: https://sigmashake.com/compare
## MCP Integration
Add to your MCP client config:
```json
{"mcpServers": {"sigmashake": {"url": "https://mcp.sigmashake.com/sse"}}}
```
## Key Concepts
- Rules: Declarative IF/THEN conditions in .rules DSL files
- Decisions: ALLOW, DENY (runtime: block), ASK, FORCE, LOG, SHADOW
- FORCE: refuses the original risky action and returns a safe substitute or required safer path for the agent to use next
- Targets: execution, read, write, search, agent, network
- Fields: command, path, content, tool, input.<key>
- Evaluation: Priority-sorted, first-match-wins; design target under 2 ms. The only published measurement is the 2026-06-10 daemon-socket run (ssg 0.29.156, n=500): p50 0.39 ms, p99 1.91 ms
## Pricing
- Simulated demo with mock data in the browser (/try, showcase); it does not run the real evaluator. Desktop/ssg is free to download and install; $5/month to use (governance eval, dashboard, hub rule bodies, metering). No annual plan, no free tier, no trial.
- Pro ($5/mo, single flat plan billed monthly): Unlimited SSG rule evaluations (policy checks, not AI model usage), private rulesets (install from private GitHub repos or publish privately to the Hub for your team), AI rule generation, priority email support
- Enterprise (custom): Unlimited SSG rule evaluations (policy checks, not AI model usage), cloud audit logs (read and export), team sharing, SSO (Okta today; SAML/Entra on request), dedicated support, on-prem deployment
- Enterprise pilots are scoped after a sales call; public collateral is not an anonymous product trial or access path.
- Default rule evaluation runs locally. If the opt-in Jev assist (bring-your-own TypeSafe key) is enabled, ssg eval sends the tool name and input to api.typesafe.ai. Running the evaluator requires the $5/month subscription, which also adds AI features and private rulesets (BYO GitHub repos or Hub-hosted team rulesets)
## Documentation
- Full docs: https://docs.sigmashake.com
- Getting started: https://docs.sigmashake.com/getting-started
- Rule syntax: https://docs.sigmashake.com/rule-syntax
- CLI reference: https://docs.sigmashake.com/cli
- Architecture: https://docs.sigmashake.com/architecture
- MCP integration: https://docs.sigmashake.com/mcp
- Claude Code setup: https://docs.sigmashake.com/claude-code-integration
## Installing as an AI agent
- Agent Install Guide (plain markdown, zero JS, curl-fetchable): https://install.sigmashake.com/agent
- Same guide on the docs site: https://docs.sigmashake.com/agent-install-guide
- Machine-parseable install pointer: https://sigmashake.com/llms-install.md