# agentsh

> Execution-layer security for AI agents — intercept file, network, process, and database activity at runtime; allow, deny, approve, or steer with full audit logs.

agentsh is an agent shell for AI agents and an execution-layer security model. Instead of relying on prompt engineering or model-side guardrails, it enforces deterministic, OS-level policy on what agent processes can read, write, execute, send over the network, or query through protected Postgres-family database services — with a full audit trail. The site documents the threat model, policy language, sandbox and MCP integrations, database proxying, observability pipeline, and the `@agentsh/secure-sandbox` (TypeScript) / `agentsh-secure-sandbox` (Python) SDKs.

The full documentation is also available as a single Markdown file at [/llms-full.md](https://www.agentsh.org/llms-full.md) for one-shot LLM ingestion.

## Overview
- [Homepage](https://www.agentsh.org/index.md): project overview, quick install, what agentsh enforces and why
- [Agent Shell](https://www.agentsh.org/agent-shell/index.md): canonical explanation for agent shell, agentshell, and agent-shell terminology
- [Glossary](https://www.agentsh.org/docs/glossary/index.md): definitions of execution-layer security, deterministic enforcement, and core terms

## Documentation
- [Setup & Integration](https://www.agentsh.org/docs/setup/index.md): platform support, container setup, AI assistant integration
- [Features](https://www.agentsh.org/docs/features/index.md): traffic steering, database proxying, Tor access control, data protection, deterministic enforcement, operations
- [Database Proxy](https://www.agentsh.org/docs/database-proxy/index.md): Postgres-family service proxying, database rules, unavoidability controls, and audit events
- [Policy Reference](https://www.agentsh.org/docs/policy-reference/index.md): file, network, command, signal, database, registry, and Tor access-control rule schema; resource limits; starter packs
- [Secure Sandbox SDK](https://www.agentsh.org/docs/secure-sandbox/index.md): adapter reference, package install security, API for the TypeScript and Python SDKs
- [MCP Security](https://www.agentsh.org/docs/mcp-security/index.md): deterministic attack detection, policy enforcement, and audit trails for MCP tool calls
- [Secrets](https://www.agentsh.org/docs/secrets/index.md): HashiCorp Vault, AWS Secrets Manager, GCP Secret Manager, Azure Key Vault, 1Password Connect, OS Keyring as secret sources
- [Observability](https://www.agentsh.org/docs/observability/index.md): session reports, OpenTelemetry export, audit log integrity

## Optional
- [AGENTS.md](https://www.agentsh.org/AGENTS.md): skill file for AI coding assistants — install, configure, and use agentsh
- [Sitemap](https://www.agentsh.org/sitemap.md): Markdown sitemap with all pages and Markdown mirrors