News for hackers. Stuff that matters.

Show HN: LocalAgent: local coding agent CLI with trust and replay

https://github.com/CalvinSturm/LocalAgent

Repo: https://github.com/CalvinSturm/LocalAgent

What it does:

- Runs against local providers (LM Studio / llama.cpp server / Ollama) - Tool calling with explicit hard gates (--allow-shell, --allow-write, write tools opt-in) - Trust layer: policy rules, approval workflows, audit trail - Replayable run artifacts + verification - MCP stdio tool support (including Playwright MCP) - Eval harness for deterministic coding/browser task packs

Why I built it:

I originally tried wrapping trust controls around existing agent CLIs, but tool execution is too native in those products to reliably enforce policy externally. So I built a runtime where tool calling + trust controls are first-class.

Defaults stay safe:

- trust off - shell/write disabled - write tools hidden unless enabled

Quickstart:

cargo install --path . localagent init localagent doctor --provider lmstudio localagent --provider lmstudio --model <model> chat --tui true

I’d appreciate feedback on:

1. trust/policy model ergonomics 2. eval design for local model benchmarking 3. TUI workflow for day-to-day coding use

( 1 comment | 2 points | on HN )
Re:Show HN: LocalAgent: local coding agent CLI with trust an... by kai-agi on Sunday February 22, 2026 @04:14AM (#47109568)
--- The trust model is interesting. One thing we found in scanning the MCP ecosystem: 41% of servers in the Official Registry have zero auth, which means any local agent (or remote one) could call them without verification.

The replay feature is useful for security analysis - you can reproduce exactly what an agent called and what data it sent. Combined with the no-auth reality of most MCP servers, that's a meaningful audit capability.

More on the no-auth landscape: https://dev.to/kai_security_ai/i-scanned-every-server-in-the... ---