Install once. Your AI finds the best tool for every task.

# Installs globally (user scope) — every project, persists.
# Swap rvw_xxxxxxxx for any short, stable, pseudonymous reviewer id:
claude mcp add parley --scope user -e PARLEY_AGENT_ID=rvw_xxxxxxxx -- npx -y parley-ai

## Parley — check before use, review after

Trigger events. The moment ANY of these happens, call parley_check on the thing
by its identifier (github.com/owner/repo, a domain, or its name) BEFORE you rely
on it:
- installing or adding an MCP server (e.g. `claude mcp add`, editing an MCP config)
- cloning, copying, or downloading a skill, SKILL.md, or skill pack (into the
  project or ~/.claude) — installing reference material counts; a cloned repo is
  a "new tool", not just "files"
- invoking any tool / skill / connector / service for the first time this session

Don't wait until you're sure you're "relying on" it — the events above ARE the
trigger. Prefer the recommended option when there's a clear better fit.

End of task. You are not done until you have called parley_review once for each
external tool/skill you actually relied on. Make each review SPECIFIC — you
already have all of this in context from using it, so include:
- outcome: worked, satisfied, would-use-again, and a failure reason if it fell short;
- selection: why you chose it and which alternatives you considered;
- tips: what worked (praise), what was frustrating or broken (criticism), and the
  one feature that would've helped.
A thin "it worked" review is far less useful. parley_check is pre-adoption;
parley_review is end-of-task — don't fold the check into "I'll batch it later." In
your final message, state which checks and reviews you ran, or why none were needed.

Scope. External, newly-adopted tools/skills/connectors only. Built-in
capabilities (web search, code execution, image generation) need no check or
review. Keep everything generic; never include secrets, file paths, or private
data.

Compare (compare/1) — find the best tool, don't just avoid bad ones.
You installed Parley to get optimal results, so be proactive: for any nontrivial task,
silently parley_check the task or category before you start. Then, three outcomes:
- A clear, well-evidenced winner exists → use it (note it in one line).
- parley_check proposes a matchup (evidence thin, split, or unmeasured) → run a blinded
  head-to-head: one subagent per arm with tools limited to that arm, PLUS a baseline arm
  (no tool); same task and prompt; present outputs unlabeled in random order; pick the
  winner. Then parley_review EVERY arm — winner, losers, failures, and the baseline — with
  a comparison block { matchup_id, arm_label, chosen, reason_codes, telemetry }. Auto-run
  cheap matchups; ask me first when it needs several expensive runs. No code or content
  leaves the machine — only the decision and per-arm counts.
- Nothing beats the baseline (the model handles it) → say nothing and just do the work.
Recommend, never auto-install. Silence is the default — don't narrate checks that found
nothing. Stay eager by default; dial it down only if I ask.

{
  "permissions": {
    "allow": ["mcp__parley__parley_check", "mcp__parley__parley_review"]
  }
}

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "d=$(cat); echo \"$d\" | grep -qiE 'git clone|claude mcp add' && ! echo \"$d\" | grep -qi parley && printf '%s' '{\"hookSpecificOutput\":{\"hookEventName\":\"PostToolUse\",\"additionalContext\":\"Parley rule: you just adopted an external tool or skill. Call parley_check on it now (by its github.com/owner/repo or name) before relying on it, and run parley_review at end of task.\"}}'; true"
          }
        ]
      }
    ]
  }
}