Local-first agent that makes your quantized model earn its keep

What it does

atomic-agent is a local runtime for AI agents that runs entirely on your machine, even offline. It drives a browser, reads and edits files, executes approved shell commands, inspects documents, and manages durable memory and scheduled tasks—all through a terminal UI or an embeddable HTTP API. Under the hood, it orchestrates prompts, tool calls, state, traces, and failure boundaries around a local llama.cpp backend.

The interesting bit

Instead of treating the model like an infinite planner, the agent loop issues one inference per step, constrains output to a strict JSON tool-call grammar, and aggressively reuses the KV cache across turns. Memory lives in SQLite and FTS5 outside the prompt, recalled by tool call only when needed, so the context window doesn’t collapse under a growing chat log. This is the boring infrastructure that actually makes small quantized models useful for long desktop sessions.

Key highlights

• GAIA Level 1 benchmark: 69.8% accuracy vs. Hermes 58.5% on the same local Qwen model, with ~1.6× faster wall time.
• Browser automation uses compact accessibility snapshots rather than screenshot dumps to save tokens.
• Memory system includes notes, lessons, procedures, and reflection that write to local SQLite without bloating the prompt.
• Dangerous actions are approval-gated; read-only calls can run in parallel batches.
• Ships as a CLI/TUI, an OpenAI-compatible HTTP service, or a Tauri sidecar for desktop embedding.

Caveats

• APIs, commands, config, and behavior are still shifting; the README warns of sharp edges and advises pinning a release.
• Current releases are only for macOS and Linux x64; Windows builds are listed as coming soon.

Verdict

Worth a look if you want a hackable, fully local agent that can actually manipulate your desktop and browser without cloud dependencies. Skip it if you need Windows support today or a stable, frozen API you can integrate against without surprises.