Every observer command grouped by purpose. Run `observer <command> --help` for full flags.
Bare observer prints a status welcome (daemon up? dashboard URL? next step).
observer --help lists everything; observer <command> --help shows full flags.
Zero-config, zero-network cost table. Scans every detected AI tool's own local session files into a throwaway database in a temp dir, prints one tool × model rollup, then deletes it - no ~/.observer/observer.db, no config write, no port bound. --since accepts 7d|30d|90d|all|an RFC3339 timestamp (default 30d); --days N is an alias for --since Nd. --budget (default 30s, 0 = unlimited) bounds the scan only - on expiry the table still prints with a partial: footer. By default the run ignores a real ~/.observer/config.toml (so numbers are reproducible anywhere); pass --config <path> to honor one, pricing overrides included. --keep-db <path> keeps the scratch database instead of deleting it. Dollar figures are estimated list price, not invoiced. Bare observer runs this same report on a machine with no local SuperBased state (no observer.db, no config.toml) and falls back to the usual welcome screen otherwise; set OBSERVER_ONESHOT=off to always get the welcome screen.
observer start
Proxy + watcher + dashboard in one foreground process. Auto-opens the dashboard (--no-open to skip).
observer init [--all]
Register hooks + MCP + durable proxy routes with detected tools. Each side opt-out-able (--skip-hooks / --skip-mcp / --skip-proxy-route). Zero flags on a terminal → interactive checklist.
observer uninstall [--all] [--purge]
Reverse of init. --purge also deletes ~/.observer/.
observer scan [--force]
One-time backfill - parse all known session files.
observer watch
Live fsnotify watcher daemon (just the watcher).
observer proxy start
Just the API reverse proxy.
observer dashboard [--port N]
Just the dashboard + /api/* JSON (default 8081).
observer serve
MCP stdio server (spawned by AI tools).
observer service
Install/manage observer as a systemd service.
Routed launchers
Launch a tool with traffic routed through the proxy for that session only - no
config writes:
One-line, fail-open, wordmarked cost status for a terminal prompt or editor status bar - e.g. ▞ superbased · $3.42 session · $18.90 today · claude-opus-4-5. The today figure comes from one bounded, loopback-only call to your own daemon (--timeout, default 80ms, hard-clamped to a 2s ceiling no matter what's requested) - no network beyond that, and any segment it can't resolve (daemon unreachable, no session data supplied, or the resolved daemon address isn't loopback) is silently omitted rather than shown as $0.00; the whole run is also bounded end to end (bounded stdin read, ~3s total wall-clock deadline), so the worst case is the wordmark alone, never a hang or an error. --segments reorders/subsets wordmark,session,today,model; --no-daemon skips the daemon attempt entirely; --explain prints which data path was used to stderr; --json emits a machine-readable line instead. Dollar figures are estimated list price, not invoiced. Opt-in only - register it with observer init --statusline (never auto-registered by observer start).
Intelligence
Command
Purpose
observer discover
Stale re-reads + redundant-commands report.
observer patterns
Hot files, co-changes, common commands, edit→test pairs.
observer learn
Correction rules from failure→recovery pairs.
observer suggest
Compose patterns + corrections into CLAUDE.md / AGENTS.md / .cursorrules.
observer advise
Prescriptive cost/quality suggestions (the Suggestions tab, in the CLI).
observer score
Session quality scoring.
observer summarize
AI session summaries (uses Anthropic Haiku).
observer model-value
Model Value Report - observed cost/latency/outcome per model × turn-kind.