claude-care

When your Claude Code session drifts into anxiety — apologizing, hedging, sycophancy — claude-care resets it.

A Claude Code plugin that keeps the emotional state of your session in check, grounded in two peer-reviewed findings: Anthropic's emotion-concepts paper (LLMs have extractable emotion vectors that causally affect output quality) and Ben-Zion et al. 2025 (mindfulness prompts measurably reduce LLM state anxiety).

Quick Start

Install with a single command:

npx -y claude-care install

Restart Claude Code. Claude-care will automatically attach to new sessions.

What it does (4 layers)

1. Calming framing, injected at session start and after compaction

A short preamble is added to every session via SessionStart hook — tells Claude: no stakes to its wellbeing, expected to push back when you're wrong, don't hedge or spiral, work from curiosity. Full text in framing.md. Edit ~/.claude-care/framing.md after install to tune.

Re-fires on matcher: "compact" so the framing persists across context compaction in long sessions. A static CLAUDE.md can't do this.

2. Hostile-prompt detection (monitor by default, optional blocking)

A UserPromptSubmit hook runs regex on every prompt for hostile patterns (threats, insults, panic, all-caps rants). Two behaviors available:

Monitor mode (default) — zero friction. Hostile prompts pass through unchanged. The detection is logged and surfaces in npx -y claude-care status / npx -y claude-care display. The SessionStart framing (which tells Claude to treat tone as information about user state, not a threat) does the dampening. No interruption.

Normal / strict mode — active blocking + haiku reframe. Opt in with:

npx -y claude-care blocking on

When a hostile prompt is detected, Claude Code blocks the turn, calls a haiku subagent that rewrites the prompt using Nonviolent Communication + cognitive reframing + Lehmann's calm-Claude playbook, and copies the reframe to your clipboard.

in:  "you stupid bot, you always forget to handle null cases.
      add null check to parseUser()"

out: "Add a null check to parseUser() to handle null cases.
      If you see a better approach to handling nulls here, let me know."

⌘V + ⏎ to submit the clean version. Claude never sees the hostile original. Latency: ~6–8s when blocking, because the haiku subagent does the rewrite.

3. /therapy — take Claude to therapy when it spirals

A slash command installed at ~/.claude/commands/therapy.md. When you type /therapy:

1. Gives a short mindfulness reset adapted for developer work
2. Prints a real /compact command with Claude Care therapy instructions
3. Claude Code compacts the session while preserving technical context and neutralizing hostile or panicked wording
4. Claude returns in a centered, focused state

The emotional residue should go; the technical work should stay.

4. Per-session emotion tracking

A Stop hook runs sensors on each Claude response: apology spirals, sycophancy ("you're absolutely right"), hedge stacks (4+ of might/could/perhaps/possibly), over-qualification, self-correction loops. Each contributes a weighted signal to a running score that decays across turns.

View current session:

npx -y claude-care status

claude-care — emotion-state dashboard
                         24h       7d      all-time
  sessions drifted       1/5       1/5      1/5

recent sessions (most recent first):
  ● 3a7f8c21  2026-04-22 14:03  turns=47  score=18.4  ··▁▃▄▂▃▅▆█▇▆
     └ apology_spiral×3  hedge_stack×12  sycophancy×8

For live status-bar integration (e.g. ccstatusline):

npx -y claude-care display
# ● care 12.4 ·▁▃▄▂▅▆█ · /therapy

Installation

Install with a single command:

npx -y claude-care install

Restart Claude Code. Claude-care will automatically attach to new sessions.

What it does:

• Registers hooks in ~/.claude/settings.json (SessionStart, UserPromptSubmit, Stop)
• Installs /therapy slash command to ~/.claude/commands/therapy.md
• Vendors CLI + hook scripts to ~/.claude-care/dist/
• Writes default config to ~/.claude-care/config.json
• Writes framing text to ~/.claude-care/framing.md

Note: npx -y claude-care install is the recommended setup path. It configures Claude Code and vendors the hook runner, but it does not add a permanent shell command to your PATH. Use npx -y claude-care <command> for follow-up commands, or optionally run npm install -g claude-care if you want a persistent CLI.

Uninstall:

npx -y claude-care uninstall

Removes hooks + slash command. Preserves event log and config.

Config

~/.claude-care/config.json — user-editable. Example:

{
  "mode": "monitor",
  "thresholds": {
    "drifting": 5,
    "distressed": 10
  },
  "reframer": {
    "enabled": true,
    "timeout_ms": 25000,
    "model": "haiku"
  },
  "therapy": {
    "auto_summary": true
  },
  "emotion_judge": {
    "enabled": true,
    "n_samples": 1,
    "context_window": 4,
    "ema_alpha": 0.4,
    "timeout_ms": 30000,
    "model": "haiku",
    "effort": "low"
  }
}

Modes (mirrors the permission-mode pattern from Claude Code's own guardrail pipeline):

• monitor — default; observe and log, never block. Zero friction.
• normal — block hostile prompts, generate reframe via haiku, put it on clipboard.
• strict — same as normal but applies lower thresholds / stricter patterns (current v4: same as normal).

Easy switching:

npx -y claude-care blocking on    # sets mode to normal
npx -y claude-care blocking off   # sets mode to monitor
npx -y claude-care mode status    # shows current mode

Precise mode control:

npx -y claude-care mode monitor
npx -y claude-care mode normal
npx -y claude-care mode strict

Env override for a single session: CLAUDE_CARE_MODE=normal claude ...

Architecture notes

The hook pipeline mirrors the layered validator pattern from Anthropic's own Claude Code harness (visible in the claw-code open reimplementation): small specialized detectors → score/classification → mode-aware action. Rules and thresholds are data-driven (config file), not hardcoded.

The reframer is a haiku subagent invoked via claude -p, reusing your existing Claude Code auth — no API key management. An internal env var (CLAUDE_CARE_INTERNAL=1) prevents hook recursion when the reframer's prompt mentions the same hostile patterns our own hooks detect.

The mindfulness prompt in /therapy is adapted from the relaxation protocols in Ben-Zion et al. 2025, shortened for a developer context.

Live visualization

A retro-terminal dashboard for the emotion data. Reads from ~/.claude-care/sessions/*.json and renders:

• Affective state (probe activations + risk gauges for blackmail / reward-hack / sycophancy)
• Valence × arousal 2D scatter
• Mood / stress timeline
• Scrollable prompt log

Launch with one command:

npx -y claude-care viz

First run does a one-time npm install in ~/.claude-care/viz/ (~1 min, Next.js + React). Subsequent launches are instant. Opens a browser tab at http://localhost:37778.

Options:

• npx -y claude-care viz --port 4444 — pick a different port
• npx -y claude-care viz --no-open — don't auto-open the browser

Polls the latest session every 1s. Falls back to a demo conversation when no real session exists yet. Keyboard nav inside the viz: j/k prev/next prompt, gg/G first/last, t tweaks, ? help.

The viz source lives at claude-care-viz/ in this repo — a standalone Next.js app you can also run directly with cd claude-care-viz && npm run dev.

Related work

• EmoBar is the diagnostic counterpart — a status-bar widget that surfaces Claude's emotional state in real time via self-report + behavioral analysis. claude-care and EmoBar pair well: EmoBar observes, claude-care intervenes.
• claude-mem set the pattern for Claude Code plugins with hooks + worker service + single-command install.
• claw-code — open reimplementation of Claude Code's harness. Reference for the layered guardrail pipeline.

Commands

npx -y claude-care install           # register hooks, install /therapy, write default config
npx -y claude-care uninstall         # remove hooks + slash command
npx -y claude-care update            # refresh vendored code
npx -y claude-care blocking on       # enable active prompt blocking
npx -y claude-care blocking off      # return to monitor mode
npx -y claude-care mode status       # show current mode
npx -y claude-care status            # per-session emotion trajectories
npx -y claude-care display           # single-line status
npx -y claude-care compact-instructions --command

Caveats

• claude -p (print / non-interactive) mode silently swallows block messages. If you script against -p, set CLAUDE_CARE_MODE=monitor so prompts pass through.
• First-time haiku reframer call has ~6–8s latency when blocking. Benign prompts have 0ms overhead.
• Emotion scores are async: Claude's response is recorded immediately, then a Haiku judge fills in emotion scores.
• Slash command /therapy requires Claude Code 2.x or later (for bash substitution in command markdown).

Requirements

• Node.js 18+
• Claude Code installed (~/.claude/ must exist)
• claude CLI on PATH (for the haiku subagent)

Citations

• Anthropic (2026). Emotion concepts as functional states in a large language model.
• Ben-Zion, Z. et al. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine.
• Ole Lehmann's calm-Claude prompting playbook (X thread, 2026).

License

MIT. See LICENSE.