#84 docs: refresh the README for the last three weeks of changes

[Anonymous]

Originally created by: sosidudku1

Problem

The README drifted three weeks behind main and carried an untrue data claim: "Your data never leaves" / "Nothing leaves unless you configure it to", while anonymous usage analytics and crash reporting are on by default.

Change

Honesty

• "Why Local-First" now states the verifiable fact: all state lives under <stateDir> as plain files and SQLite databases, with a pointer to Privacy and Egress.
• "Privacy and Egress" documents the default-on anonymous analytics (provider and model names, random install id, turn-shape numbers; never message content, file paths, or tool arguments), Sentry crash reports on the same switch, and both off ramps: /privacy analytics off and analytics.enabled: false.
• The egress list drops "exactly" and adds the missing sources: web search, ClawHub skill installs, the startup update check, analytics and crash reporting.
• "keeps quality high" softened to "keeps quality usable" in both places.

Fact fixes

• Build targets: macOS (Apple Silicon), Linux x64 / arm64, Windows x64; managed local models on Linux are x64 only, arm64 uses external mode.
• The serve route list drops traces (no such route) and adds skills.
• Web search providers: Exa, DuckDuckGo, Brave, SearXNG.
• Image alt texts now spell "Atomic Agent".

Last Three Weeks of Features

• TUI: a handy slash commands paragraph (/help, /tools, /model, /privacy) and PgUp/PgDn chat scrolling.
• Cloud providers: full live model catalogs instead of a short hardcoded list, own /v1/models for OpenAI-compatible servers, mid-session switching with /model.
• External llama-server: the LLM tab External pane and the honest health probe (a 503 while the model loads is loading, not dead; foreign OpenAI-compatible runners are recognized).
• Managed models: the daemon stops when the last session exits; localModels.managed.stopOnExit: false keeps the model warm.
• Telegram: the live progress bubble (silent, step labels only) and its telegram.progressIndicator switch.
• Reasoning models: reasoning-only completions are recovered instead of failing the turn.

Small Accuracy Touches

• SSRF guard noted on raw HTTP requests too, the ARIA snapshot character budget, MCP stderr surfaced on connect failures, and the automatic GPU backend pick (Metal / CUDA with driver 610+ headers / Vulkan / CPU fallback).

New

• A "Talk to Us" section after Quick Install: book a call with the team at cal.com/atomicagent/demo.

Benchmarks, GAIA numbers, TurboQuant claims, the demo GIF, and the install commands are untouched. The model picker filter from [#83] is deliberately not described yet; that line lands together with [#83].

[Anonymous]

Ticket changed by: Ooooze

• status: open --> closed