superview

The terminal observatory for Claude Code — observability for AI-assisted development. Browse every workspace, session, and behavioral signal across all your Claude Code work, locally.

Why superview

AI coding agents went mainstream in 2024–25. Engineering teams now ship significant fractions of their code through tools like Claude Code, but they have no visibility into how those agents actually behave: where they thrash, where they loop, where users get frustrated, where token spend goes off the rails.

superview is the observability layer for that work. It reads Claude Code's session transcripts on your machine and surfaces:

1. What happened — every workspace, session, conversation, and tool call.
2. What went wrong — automated detection of behavioral anti-patterns (edit thrashing, error loops, sentiment drift, abandonment, restart clusters, and more).
3. Where the spend is — token usage rolled up by workspace, session, and time range.

Read-only, runs locally, no account required.

Quick start

# Try it without installing
npx superview

# Or install globally
npm install -g superview
superview

The CLI is also available as sv and coverview — all the same binary.

sv               # short alias
sv --help        # full command reference
sv --version

At a glance

What it does

Claude Code writes every conversation, tool call, error, and interrupt to structured JSONL files in ~/.claude/projects/. superview reads those files locally and gives you:

• Workspace list — all projects Claude Code has touched, sorted by recent activity.
• Session browser — pick a session and read the conversation or browse raw events.
• Aggregate stats — message counts, tool call rates, error rates, session durations.
• Signal analysis — automated detection of behavioral anti-patterns:
  • Edit thrashing (same file edited 5+ times)
  • Error loops (consecutive tool failures)
  • Excessive exploration (high read-to-edit ratio)
  • Correction-heavy sessions (user repeatedly saying "no", "wrong")
  • Keep-going loops (user nudging Claude because it stops early)
  • Repeated instructions (Claude didn't act on something the first time)
  • Negative sentiment drift (conversation tone worsens over a session)
  • Rapid corrections (user responds in under 10 seconds)
  • High abandonment rate (many very short sessions)
  • Restart clusters (multiple sessions started within 30 minutes)

Your transcripts stay on your machine. superview reads only ~/.claude/projects/ and never transmits prompt or file content. See PRIVACY.md for the full data inventory.

How superview is different

| | Raw ~/.claude/projects/ JSONL | Anthropic Console | superview | | -------------------------------- | ------------------------------- | --------------------- | ---------------------------------- | | Cross-session pattern detection | No | No | Yes (10+ signal detectors) | | Cost / token rollups across runs | DIY | Per-account, lossy | Per-workspace, per-session | | Local-first, offline | Yes | No | Yes | | No account / sign-up | Yes | No | Yes | | Multi-provider roadmap | n/a | No | Yes (Cursor, Cline, Aider planned) |

Requirements

• Node.js 18+
• Claude Code installed (provides the session transcripts at ~/.claude/projects/)

Navigation

Workspace list
  arrow keys / j k     navigate
  Enter                select workspace
  type # + Enter       jump to item by number
  q                    quit

Session browser
  arrow keys / j k     navigate
  Enter                open session
  q                    back

Conversation reader (inside a session)
  arrow keys / j k     scroll line by line
  PgUp / PgDn          scroll fast
  g / G                jump to top / bottom
  t                    toggle thinking blocks
  q                    back

Event pager (inside a session)
  arrow keys / j k     navigate events
  left / right or p n  previous / next page
  Enter                open focused event detail
  r                    switch to conversation reader
  f                    show files touched in session
  e                    toggle errors-only view
  s                    session summary
  type # + Enter       jump to event number
  q                    back

Signal severity

| Badge | Meaning | |----------|------------------------------------------------| | [CRIT] | Severe pattern, likely causing major friction | | [HIGH] | Significant pattern worth investigating | | [MED ] | Moderate signal, monitor if it recurs | | [LOW ] | Mild signal, context-dependent |

Privacy and telemetry

superview sends anonymous usage events (event names, aggregate counts, a random installation ID) to PostHog by default. Prompts, file paths, and session content are never transmitted. The full data inventory and opt-out instructions are in PRIVACY.md.

To opt out:

echo '{"anonymousTelemetry": false}' > ~/.superview/config.json

What's next

• The CLI is and will remain MIT-licensed open source.
• A hosted multi-user version with team dashboards, SSO, and audit logs is in private beta — see docs/cloud.md and the waitlist link there.

Development

npm install        # deps
npm run build      # one-shot build
npm run dev        # watch mode
npm test           # unit tests
node dist/cli.js   # run the built CLI

The entry point is src/cli.ts. Signal detectors live in src/signals/. The indexer that scans ~/.claude/projects/ is src/indexer.ts. Architecture deep-dive: docs/ARCHITECTURE.md. Contributor onboarding: CONTRIBUTING.md.

Documentation

• PRIVACY.md — what telemetry collects and how to opt out
• SECURITY.md — how to report vulnerabilities
• CONTRIBUTING.md — local setup, signals, PR workflow
• CHANGELOG.md — release notes
• docs/ — Claude Code data model background and architecture docs
• docs/cloud.md — hosted/team version roadmap

License

MIT — see LICENSE.