Show HN: Locki – AI sandboxing for real-world projects
github.comHello HN! I was bothered that no existing sandboxing solution can run my $job project, due to the limitations of running inside an OCI container, or some sort of limited process like landlock / bubblewrap etc. My options were spawning a new VM per worktree, which takes minutes to boot and allocates a chunk of RAM... or build a custom solution which uses just a single VM plus LXC containers -- the kind of containers with a full init system, capable of running systemd services, OCI containers and even full Kubernetes distributions.
So here it is. Fresh container start takes <10s. Works best with VSCode, which will neatly show changes from all worktrees in the sidebar, letting you review, edit and commit them easily. Let me know what you think! hey this is great idea. Can you share what kind of $job project you're working on? Actually I can, it's open source: https://github.com/kagenti/adk It's an agent development platform. It needs MicroShift, which is notoriously not easy to run locally -- ADK makes it simple by using a prebuilt Lima VM for CLI deploy, but that doesn't work well in development. Noice! Agent deployment infra is exactly the layer that's missing right now. Everyone's building agents but nobody's solving the "get it into production reliably" problem.
The MCP connector support is interesting to me, I'm working on an MCP server that handles payment injection for agents (card number never enters LLM context, injected via CDP at the browser layer). Biggest pain point is testing against real deployment setups vs my janky local Chrome + CDP dev loop.
Curious: when you add an MCP server to an ADK agent, does the agent discover tools at startup or can you reconfigure at runtime? Thinking about cases where you'd want to swap a mock payment tool for a live one during staging → prod promotion.
Going to try ADK this week — been needing something exactly like this for integration testing. To complicate it further, ADK is quite "early" in development, but it's a fusion-fork of two more mature projects: Red Hat's Kagenti, and Linux Foundation's (orig. IBM) Agent Stack (https://github.com/i-am-bee/agentstack/). ADK is basically Agent Stack but with the custom backend replaced by Kagenti. Regarding MCP: the current idea is dependency injection using A2A extensions: agent says what it wants (LLM endpoint, MCP server, vector database, etc.) and the caller provides. Works great for direct communication, gets weird with subagents -- we may iterate on it a bit more.