Anthropic’s latest Claude model, running twice in the background of your Mac

For general availability, the two current releases are Claude Sonnet 4.6 (the everyday default, fast and cheap enough to run continuously) and Claude Opus 4.7 (the stronger, pricier ceiling, GA since April 22, 2026 at $5 input / $25 output per million tokens). A third release, Claude Mythos (codename Capybara), was unveiled behind Project Glasswing and is not on any public plan. Inside Fazm, the main chat, the floating bar, and the background observer all run against claude-sonnet-4-6 by default, with Opus 4.7 available per-turn in the Smart slot of the model picker.

Because a consumer app cannot afford to ask the model to do two jobs at once. Fazm calls acpBridge.warmupSession in Desktop/Sources/Providers/ChatProvider.swift at lines 1047 to 1051 with three session configs: main, floating, and observer. Each one is its own Claude session with its own system prompt. The main session handles long-form chat in the docked window. The floating session drives the always-on-top command bar. The observer session runs silently, watches conversation turns flow through the other two, and persists what matters to a user-local memory directory without ever touching the UI.

The Chat Observer is a second Claude Sonnet 4.6 session that Fazm spawns in parallel with the user-facing chat. Every user turn and assistant turn flowing through the main or floating session is also pushed onto an observer buffer via bufferChatObserverTurn (acp-bridge/src/index.ts line 1158). When the buffer reaches CHAT_OBSERVER_BATCH_SIZE, which is exactly 10 turn pairs (20 messages), flushChatObserverBatch fires a single session/prompt request against the observer session. The observer's job is to read MEMORY.md, decide what is genuinely new, write topic files using Claude's built-in memory tools, and call save_observer_card to surface one-line notes to the user.

Three reasons. First, token cost: running Sonnet 4.6 on every single turn would double your per-message bill for a background job that does not need that much resolution. Second, signal: a single turn is rarely meaningful, but 10 pairs is enough context for the model to tell 'the user keeps asking about CRM workflows' from 'the user was debugging a typo'. Third, concurrency: the observer runs through a single at-a-time gate (chatObserverRunning flag in index.ts line 1168), so batching prevents overlap when the user is typing fast.

The observer is deliberately stripped down. In acp-bridge/src/index.ts line 1023, the buildMcpServers function returns early for sessionKey === 'observer' after attaching only fazm_tools. No Playwright. No macos-use. No WhatsApp. No browser overlay. The observer literally cannot click, type, browse, or take a screenshot. Its tool surface is reduced to file system, SQL against the local app DB, save_observer_card, and the browser-profile query/edit pair. That is the entire power budget for what a second Claude session is allowed to do in the background.

The system prompt in Desktop/Sources/Chat/ChatPrompts.swift line 566 instructs the observer: when you detect a repeated workflow three or more times, draft a skill. The observer reads chat_messages from the local SQLite database, checks ~/.claude/skills for what already exists, and if a genuine repeat emerges, writes a new SKILL.md file and calls save_observer_card with type 'skill_created'. The user sees a card in the floating bar saying 'Created skill: {name}' and can dismiss it to undo.

Yes, so Fazm keeps it entirely local. The observer's memory directory is on the user's disk, not a Fazm-hosted database. The observer's allowed tools do not include any network calls back to Fazm servers beyond the normal anonymized analytics. The fazm_tools MCP server runs as a stdio subprocess connecting to the Swift app via a Unix socket (FAZM_BRIDGE_PIPE), so the observer's reads and writes never leave the machine. The one thing it does call Anthropic for is the model inference itself, which is the same call path as any other Claude API turn.

Not that we have found on Mac. Most consumer AI apps use one LLM call per user message, period. Some frameworks expose background agents, but those are developer tools, not consumer products. The closest architectural cousin is Claude Code itself, which ships its own memory files under a single session. Fazm's twist is running the memory pass as a separate Sonnet 4.6 session so the main chat's context window stays clean and the observer can be given a different system prompt and a smaller tool surface.

The observer picks it up the same way the main chat does. The ACP bridge reports availableModels on every session/new, Fazm's ShortcutSettings.updateModels substring-matches the ID against haiku / sonnet / opus, and the observer's warmup config is swapped to whatever the family map currently resolves for Sonnet. There is no build, no app update, and no user action. If Anthropic renames the family entirely, the observer falls back to the raw model name from the API, sorted to the end of the picker, until a new build ships.

Yes. The observer session is created during ACP warmup. Users who do not want a parallel memory pass can disable the feature in settings, which clears the observer session and stops bufferChatObserverTurn from being called after each main/floating turn (acp-bridge/src/index.ts lines 1748 to 1752). The main chat keeps working identically. The only thing you lose is the automatic MEMORY.md and the auto-drafted skills.