agentcage
agentcage is an experimental open-source defense-in-depth security sandbox for autonomous AI coding agents. It confines the workload to an internal-only network and forces internet access through an inspected egress proxy that enforces policy, injects decoy-backed secrets on the wire, and records decisions.
At a Glance
- Individual developers using autonomous coding agents
- Software engineering and platform teams
- CI/CD and developer-tooling teams
- Security-conscious enterprises evaluating agent deployments
- +1 more
AI Tools by agentcage
(1)agentcage
Open Source AI Agent Sandbox
Discussions
No discussions yet
Be the first to start a discussion about agentcage
Latest News
v0.54.1 released; updates wasmtime to 49.0.5 to address three RustSec advisories and raises the source-build requirement to Rust 1.99.
v0.54.0 released with self-update support and JSON output for doctor, cage show/status, and cage list commands.
v0.53.1 released as a maintenance release.
v0.53.0 released with Claude OAuth-token authentication for autonomous decider/watcher agents, system-prefix configuration, and prebuilt-binary installation by default.
Products & Services
MIT-licensed CLI and runtime that creates and manages isolated cages for AI agents. It supports ephemeral runs and persistent cages, with rootless Podman on Linux, Apple Container microVMs on newer Apple Silicon macOS, and Lima VM isolation for other macOS/Linux configurations.
The network choke point through which cage traffic must pass. It terminates TLS with a per-cage CA, validates SNI/Host identity, applies allowlists and content inspectors, performs secret substitution/redaction, supports WebSocket inspection, and emits structured audit/HAR records.
Preconfigured hardened images and policies for Claude Code, OpenAI Codex, Pi, OpenClaw, and base environments including Ubuntu, Debian, and Arch.
Technical documentation covering installation, architecture, the eight-layer security model, policy API, traffic watcher, configuration, CLI, secrets, protocol relays, custom WebAssembly inspectors, and operational guides.
Market Position
agentcage differentiates itself from a basic container wrapper by combining network confinement, a Rust L7 egress proxy, TLS/content inspection, wire-time secret injection, multiple isolation backends, autonomous policy controls, and forensic capture. A directly comparable open-source project is Containarium, which uses SSH-native isolation, eBPF egress policy, Kubernetes/LXC backends, GPU passthrough, and an MCP-native CLI; PNNL’s Agent Cage is another related Docker-based prototype focused on TLS inspection, domain whitelisting, HTTP method filtering, payload inspection, and audit logging. agentcage’s distinctive emphasis is defense-in-depth and zero-leak secret handling at the egress boundary.
Leadership
Founders
Luca Martinetti
Project author and principal maintainer. PyPI identifies Luca Martinetti as the sole maintainer; the GitHub copyright notice covers 2025–2026. His public LinkedIn profile lists TrueLayer as his current organization and prior associations with Plaid, Upsight, and Domainsbot; his GitHub profile shows substantial authorship of agentcage.
Founding Story
The project was started to address the security risk created when autonomous coding agents simultaneously receive access to private credentials, unrestricted internet connectivity, and arbitrary code execution—the so-called lethal trifecta. Its initial vision was a portable cage in which the agent never sees real secrets, has no direct internet gateway, and can operate only through a policy-enforcing, auditable egress boundary.
Business Model
Revenue Model
The project is distributed as MIT-licensed open-source software through GitHub releases, source builds, and PyPI. The public sources describe no hosted service, subscription, paid license, or commercial pricing model.
Target Markets
- Individual developers using autonomous coding agents
- Software engineering and platform teams
- CI/CD and developer-tooling teams
- Security-conscious enterprises evaluating agent deployments
- Researchers and operators studying agent isolation and egress control
- Running autonomous coding agents against untrusted repositories, webpages, and dependency sources
- Interactive local development with Claude Code, OpenAI Codex, Pi, or OpenClaw
- CI jobs that need agent code execution while limiting network and credential exfiltration
- Persistent agent workspaces with auditable network traffic and controlled domain grants
- Security testing and forensic review of agent network behavior