Browse the personal kit: 74 agents, 94 commands, 100 skills. Call this when the user mentions Supabase (RLS, branching, migrations, Edge Functions, Custom Claims, Postgres Roles, Storage, Realtime, pgvector), multi-tenant SaaS, agentic harness, characterization tests, legacy refactor, observability (SLO, golden signals, error budgets), DDIA topics (consistency, replication lag, schema evolution), SRE (postmortems, toil, PRR), CI/CD (hermetic builds, pipelines), or any workflow that benefits from the canonical patterns. Use action=search to discover, action=get to read the full prompt/skill.

Project the kit into an IDE-specific layout (markdown references by default).

Detect and apply edits made directly in an IDE back to the canonical kit/.

List, fetch, or execute reusable workflow gates (regression, confidence, etc).

Failure dataset & replays — close the learning loop on failed agent runs.

Register this kit-mcp server into an IDE's MCP config (Claude/Cursor/Codex/Antigravity/Windsurf).

Read in-memory golden-signals metrics for this MCP server (counters + latency p50/p95/p99 per tool).

IMPORTANT for first contact: project kit/ into the host's native layout (.claude/agents/, skills/, commands/) so 74 agents become real subagent_types in the Agent tool, 100 skills get native auto-trigger via descriptions, and 94 commands appear as /slash-commands in the IDE. Idempotent — re-running is a no-op if already in sync. Run once per project on first kit-mcp contact; restart the IDE session after to load the new agents/skills/commands. After restart, call ack-restart to clear the marker.

Acknowledge that the IDE session was restarted after kit:auto-install. Removes the .kit-mcp-restart-required marker so kit:doctor stops warning. Called automatically by the harness when it detects the marker after reload, or manually by the user.

Custo Claude Code do dia corrente (USD + tokens por modelo) lendo JSONLs de ~/.claude/projects/. tz default UTC (paridade ccusage). Retorna shape canônico: total_usd, by_model, entry_count, deduped_count, skipped_entry_count, parse_error_count, unknown_models, pricing_source, pricing_staleness_days. Triggers: "quanto gastei hoje", "custo do dia", "spent today", "daily cost". Use cost-session para a sessão atual, cost-blocks para janelas de 5h, cost-phase para uma fase do framework.

Custo Claude Code de uma sessão específica (ou da sessão ativa auto-deduzida pelo arquivo JSONL mais recente com mtime < 30min). Retorna shape canônico + session_id, started_at, last_activity_at, source_file. Triggers: "custo da sessão", "session cost", "quanto essa conversa gastou", "current session usd". Use cost-today para o dia inteiro.

Custo Claude Code por janelas deslizantes de 5h com gap detection (entries separadas por >5h iniciam novo bloco — pattern ccusage). Retorna blocks[] com started_at, ended_at, total_usd, by_model, entry_count, is_active + shape canônico agregado. Triggers: "custo por bloco", "5h windows", "blocks cost", "ccusage blocks". Use cost-today para dia, cost-session para sessão.

Custo Claude Code correlacionado com uma fase do framework kit-mcp (.planning/phases/-*/). Cruza mtime de SPEC.md + completed_at de STATE.md + git log para inferir janela temporal. Retorna shape canônico + phase_id, phase_slug, correlation_confidence (high/medium/low/unknown). Diferencial vs ccusage: contexto de workflow. Triggers: "custo da fase", "quanto a fase X gastou", "phase cost".

Estima custo USD de um prompt ANTES de mandar para Claude. Heurística chars/4 com range ±30% (sem tokenizer real na v1.37.0 — debt em SKILL.md). Retorna estimated_input_tokens, estimated_output_tokens, estimated_usd, estimated_usd_range:[low,high], disclaimer. Triggers: "quanto vai custar", "estimativa de prompt", "estimate cost", "price this prompt".