Zerostack
Zerostack is a minimal coding agent written in Rust that emphasizes a small binary, low memory use, Cargo installation, provider/model selection, built-in prompts, and optional bubblewrap sandbox mode for bash command isolation.
Zerostack is a strong local-agent example because it shows that the local coding-agent wave is not only about running smaller models. The harness itself can be lightweight, sandbox-aware, prompt-switchable, and designed for low overhead.
The official GitHub repository describes Zerostack as a minimalistic coding agent written in Rust and optimized for memory footprint and performance. It documents Cargo installation, optional ACP support, optional bubblewrap sandbox mode, OpenRouter default setup, explicit provider/model selection, and built-in prompts. Hacker News discussion around the Zerostack launch provides strong demand evidence for lightweight coding agents, but official repository text should remain the factual basis.
- Experiment with a lightweight local coding-agent harness.
- Run coding-agent sessions with explicit provider and model selection.
- Use optional sandboxing to reduce command-execution risk during local agent work.
- Compare small Rust agent design against OpenCode, SmallCode, Claude Code, and other coding agents.
Zerostack can be installed with Cargo and run as an interactive session, monochrome TUI, one-shot prompt, continued session, or explicit provider/model invocation. Its README positions the project around low binary size, low memory footprint, and low idle CPU usage compared with larger coding-agent stacks.
- Install path: `cargo install zerostack`, with optional ACP support for editor integration.
- Sandbox option: bubblewrap can isolate bash commands through `--sandbox`.
- Model routing: provider and model can be set explicitly, with OpenRouter documented as the default setup path.
Zerostack gives readers a concrete example of local-agent design pressure: if the model is smaller or the machine is constrained, the harness has to be cheap, direct, and predictable. It is useful as a related entity for Local Coding Agents, Agent Harness, and comparisons against heavier JavaScript or hosted coding-agent products.
Performance claims should be read as project-provided until reproduced across different machines and repositories. The page value is still clear: Zerostack names a product direction where coding agents compete on footprint, sandboxing, prompt control, and provider flexibility, not only model benchmark scores.
The category Zerostack belongs to: agents that can run with local or self-hosted model choices.
The runtime, prompt, provider, shell, and sandbox layer that determines Zerostack behavior.
Security context for prompts, shell execution, sandboxing, and agent-side instructions.
Zerostack FAQ
Page-level questions for Zerostack.
Is Zerostack only useful for local models?+
No. Zerostack is useful for local-agent experimentation, but its documented provider/model flags mean it can route through provider endpoints as well. Its main distinction is the lightweight Rust harness and local-control posture, not a rule that every model must run on the same machine.
Why does Zerostack mention sandbox mode?+
Coding agents can run shell commands, so sandboxing matters. Zerostack documents optional bubblewrap support for isolating bash commands, which helps readers understand why local agents still need filesystem and process boundaries.