Tools

atomic-agent is an operator, not a chat box. It drives the same tools you’d use by hand. Independent read-only calls can run in parallel after a single inference, and tool calls are constrained into a strict JSON array (GBNF-grammar) so actions run instead of failing on a malformed request.

Capability surface

• Browser — navigate, click, type, inspect tabs, and read a compact browser state via playwright-core against Chrome, Edge, or another Chromium-family browser. Ordinary web operation uses accessibility / ARIA snapshots instead of screenshot-heavy page dumps.
• Filesystem & shell — read, write, patch, glob, grep, archive, hash, inspect processes, use the clipboard, send notifications, and run approved commands.
• Documents — extract text from PDF, DOCX, DOC, XLSX, RTF, ODT, PPTX, archives, and plain text — all locally.
• Git — status, log, diff, show, blame, and branch inspection.
• Vision — optional vision.describe for multimodal models with mmproj, kept outside the normal text transcript.
• Tasks — durable deferred turns, cron schedules, intervals, webhooks, and agent-created reminders.

How a step works

One inference produces one JSON array of tool calls. The runtime executes those calls, compresses the results, updates durable state, and asks the model for the next move.

The governing invariant is one LLM inference equals one agent step — the model isn’t treated as an infinite planner.