Give Agents Isolated Linux Sandboxes via MCP [Kilntainers]
MCP sandbox isolation for agents; E2B/Modal/Docker/WASM backends already exist separately.
The self-improving sandboxed and open-source AI agent. With persistent memory and scheduling.
Another open-source AI agent—self-improvement is just keyword-matched JSON rules.
Developers wanting local AI agents with browser control
OpenHands · Devika · AutoGen
MCP sandbox isolation for agents; E2B/Modal/Docker/WASM backends already exist separately.
Credential proxying keeps keys out of sandboxes, unlike Ramp Inspect.
Single Go binary: Telegram → Claude agents in isolated Docker with swarms, memory, Nix.
Agent fleets in hardened Docker with per-agent budgets—assumes agents will be compromised.
This is a practical, engineer-first sandbox that feels built for LLM workflows: five runtimes (Python/Node/Bun/Deno/Bash), streaming SSE output, warm container pools for sub-100ms latency, and security defaults like read-only rootfs, seccomp, and resource caps. The embeddable TypeScript API plus an agent 'skill' and on-the-fly package installs make it easy to plug into agent pipelines. My nitpick: it still depends on Docker as the trust boundary — I'd like clearer hardened defaults, policy/audit primitives, and documentation about residual host risks before using it to run fully hostile code.
Docker sandbox for AI agents with egress proxy and filesystem isolation—solves real runaway-agent fear.