diff --git a/.mcp.json b/.mcp.json
new file mode 100644
index 0000000..cde04ad
--- /dev/null
+++ b/.mcp.json
@@ -0,0 +1,7 @@
+{
+  "mcpServers": {
+    "chrysopedia": {
+      "url": "http://ub01:8097/mcp"
+    }
+  }
+}
diff --git a/docker-compose.yml b/docker-compose.yml
index 8600b6e..a3dd709 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -198,6 +198,41 @@ services:
       start_period: 10s
     stop_grace_period: 15s
 
+  # ── MCP Server (management tools for AI agents) ──
+  chrysopedia-mcp:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.mcp
+    container_name: chrysopedia-mcp
+    restart: unless-stopped
+    environment:
+      DATABASE_URL: postgresql://${POSTGRES_USER:-chrysopedia}:${POSTGRES_PASSWORD:-changeme}@chrysopedia-db:5432/${POSTGRES_DB:-chrysopedia}
+      REDIS_URL: redis://chrysopedia-redis:6379/0
+      QDRANT_URL: http://chrysopedia-qdrant:6333
+      API_URL: http://chrysopedia-api:8000
+      PROMPTS_PATH: /prompts
+    ports:
+      - "0.0.0.0:8097:8097"
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock:ro
+      - ./prompts:/prompts:ro
+    depends_on:
+      chrysopedia-db:
+        condition: service_healthy
+      chrysopedia-redis:
+        condition: service_healthy
+      chrysopedia-api:
+        condition: service_healthy
+    networks:
+      - chrysopedia
+    healthcheck:
+      test: ["CMD-SHELL", "curl -sf http://localhost:8097/mcp -o /dev/null || exit 1"]
+      interval: 30s
+      timeout: 5s
+      retries: 3
+      start_period: 15s
+    stop_grace_period: 15s
+
 networks:
   chrysopedia:
     driver: bridge
diff --git a/docker/Dockerfile.mcp b/docker/Dockerfile.mcp
new file mode 100644
index 0000000..2cd46ee
--- /dev/null
+++ b/docker/Dockerfile.mcp
@@ -0,0 +1,19 @@
+FROM python:3.12-slim
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl libpq-dev gcc \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY mcp_server/requirements.txt /app/requirements.txt
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY mcp_server/server.py /app/server.py
+
+EXPOSE 8097
+
+HEALTHCHECK --interval=30s --timeout=5s --retries=3 --start-period=10s \
+    CMD curl -sf http://localhost:8097/mcp -o /dev/null || exit 1
+
+CMD ["python", "server.py"]
diff --git a/mcp_server/requirements.txt b/mcp_server/requirements.txt
new file mode 100644
index 0000000..95c4fc0
--- /dev/null
+++ b/mcp_server/requirements.txt
@@ -0,0 +1,5 @@
+mcp[cli]>=1.9.0,<2.0
+psycopg2-binary>=2.9,<3.0
+redis>=5.0,<6.0
+docker>=7.0,<8.0
+uvicorn[standard]>=0.32.0,<1.0
diff --git a/mcp_server/server.py b/mcp_server/server.py
new file mode 100644
index 0000000..96947aa
--- /dev/null
+++ b/mcp_server/server.py
@@ -0,0 +1,828 @@
+"""
+Chrysopedia MCP Server — Docker-native management tools.
+
+Runs inside the Docker Compose stack with direct access to PostgreSQL,
+Redis (Celery broker), Docker socket, and the filesystem.
+
+25 tools across 6 domains: pipeline execution, infrastructure health,
+content queries, pipeline observability, embeddings, and prompts.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import subprocess
+from datetime import datetime, timezone
+
+import docker
+import psycopg2
+import psycopg2.extras
+import redis
+from mcp.server.fastmcp import FastMCP
+
+# ── Configuration ─────────────────────────────────────────────────────────────
+
+DB_URL = os.getenv(
+    "DATABASE_URL",
+    "postgresql://chrysopedia:changeme@chrysopedia-db:5432/chrysopedia",
+)
+REDIS_URL = os.getenv("REDIS_URL", "redis://chrysopedia-redis:6379/0")
+QDRANT_URL = os.getenv("QDRANT_URL", "http://chrysopedia-qdrant:6333")
+API_URL = os.getenv("API_URL", "http://chrysopedia-api:8000")
+PROMPTS_PATH = os.getenv("PROMPTS_PATH", "/prompts")
+COMPOSE_PROJECT = "xpltd_chrysopedia"
+
+mcp = FastMCP(
+    "Chrysopedia",
+    stateless_http=True,
+    json_response=True,
+    host="0.0.0.0",
+    port=8097,
+)
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+
+def _db():
+    """Return a new psycopg2 connection (sync, short-lived)."""
+    return psycopg2.connect(DB_URL, cursor_factory=psycopg2.extras.RealDictCursor)
+
+
+def _redis():
+    """Return a Redis client."""
+    return redis.from_url(REDIS_URL, decode_responses=True)
+
+
+def _docker():
+    """Return a Docker client from the mounted socket."""
+    return docker.from_env()
+
+
+def _query(sql: str, params: tuple = ()) -> list[dict]:
+    """Execute a read query and return rows as dicts."""
+    with _db() as conn:
+        with conn.cursor() as cur:
+            cur.execute(sql, params)
+            return [dict(r) for r in cur.fetchall()]
+
+
+def _query_one(sql: str, params: tuple = ()) -> dict | None:
+    rows = _query(sql, params)
+    return rows[0] if rows else None
+
+
+def _execute(sql: str, params: tuple = ()) -> int:
+    """Execute a write query and return rowcount."""
+    with _db() as conn:
+        with conn.cursor() as cur:
+            cur.execute(sql, params)
+            conn.commit()
+            return cur.rowcount
+
+
+def _json_default(obj):
+    """JSON serializer for types not handled by default."""
+    if isinstance(obj, datetime):
+        return obj.isoformat()
+    if hasattr(obj, "__str__"):
+        return str(obj)
+    raise TypeError(f"Object of type {type(obj)} is not JSON serializable")
+
+
+def _json(data) -> str:
+    return json.dumps(data, default=_json_default, indent=2)
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 1: Pipeline Execution
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def pipeline_status() -> str:
+    """Show currently processing, queued, and backlogged videos with counts."""
+    processing = _query(
+        "SELECT id, filename, processing_status, updated_at "
+        "FROM source_videos WHERE processing_status = 'processing' "
+        "ORDER BY updated_at DESC"
+    )
+    queued = _query(
+        "SELECT id, filename, processing_status, updated_at "
+        "FROM source_videos WHERE processing_status = 'queued' "
+        "ORDER BY updated_at DESC"
+    )
+    not_started = _query(
+        "SELECT id, filename, updated_at "
+        "FROM source_videos WHERE processing_status = 'not_started' "
+        "ORDER BY updated_at DESC"
+    )
+    errored = _query(
+        "SELECT id, filename, updated_at "
+        "FROM source_videos WHERE processing_status = 'error' "
+        "ORDER BY updated_at DESC"
+    )
+
+    # Check Celery queue
+    r = _redis()
+    celery_queue_len = r.llen("celery")
+
+    return _json({
+        "processing": processing,
+        "queued": queued,
+        "not_started": not_started,
+        "errored": errored,
+        "celery_queue_depth": celery_queue_len,
+        "summary": {
+            "processing": len(processing),
+            "queued": len(queued),
+            "not_started": len(not_started),
+            "errored": len(errored),
+            "celery_pending": celery_queue_len,
+        },
+    })
+
+
+@mcp.tool()
+def trigger_pipeline(video_id: str = "", filename: str = "") -> str:
+    """Trigger pipeline for a video by ID or filename substring. Provide one."""
+    if not video_id and not filename:
+        return "Error: provide either video_id or filename"
+
+    if not video_id:
+        row = _query_one(
+            "SELECT id, filename FROM source_videos "
+            "WHERE filename ILIKE %s LIMIT 1",
+            (f"%{filename}%",),
+        )
+        if not row:
+            return f"No video matching '{filename}'"
+        video_id = str(row["id"])
+        filename = row["filename"]
+    else:
+        row = _query_one("SELECT filename FROM source_videos WHERE id = %s", (video_id,))
+        filename = row["filename"] if row else "unknown"
+
+    import httpx
+
+    resp = httpx.post(
+        f"{API_URL}/api/v1/admin/pipeline/trigger",
+        json={"video_id": video_id},
+        timeout=10,
+    )
+    return _json({
+        "video_id": video_id,
+        "filename": filename,
+        "status_code": resp.status_code,
+        "response": resp.json() if resp.status_code < 400 else resp.text,
+    })
+
+
+@mcp.tool()
+def trigger_batch(
+    creator: str = "",
+    status: str = "not_started",
+    limit: int = 5,
+) -> str:
+    """Trigger pipeline for multiple videos. Filter by creator slug and/or status.
+    Default: up to 5 not_started videos."""
+    sql = "SELECT sv.id, sv.filename FROM source_videos sv"
+    conditions = []
+    params: list = []
+
+    if status:
+        conditions.append("sv.processing_status = %s")
+        params.append(status)
+    if creator:
+        sql += " JOIN creators c ON sv.creator_id = c.id"
+        conditions.append("c.slug = %s")
+        params.append(creator)
+
+    if conditions:
+        sql += " WHERE " + " AND ".join(conditions)
+    sql += " ORDER BY sv.updated_at DESC LIMIT %s"
+    params.append(limit)
+
+    videos = _query(sql, tuple(params))
+    if not videos:
+        return "No videos match the filter."
+
+    import httpx
+
+    results = []
+    for v in videos:
+        try:
+            resp = httpx.post(
+                f"{API_URL}/api/v1/admin/pipeline/trigger",
+                json={"video_id": str(v["id"])},
+                timeout=10,
+            )
+            results.append({
+                "video_id": str(v["id"]),
+                "filename": v["filename"],
+                "status": resp.status_code,
+            })
+        except Exception as e:
+            results.append({
+                "video_id": str(v["id"]),
+                "filename": v["filename"],
+                "error": str(e),
+            })
+
+    return _json({"triggered": len(results), "results": results})
+
+
+@mcp.tool()
+def revoke_pipeline(video_id: str) -> str:
+    """Revoke/cancel a pipeline run for a video."""
+    import httpx
+
+    resp = httpx.post(
+        f"{API_URL}/api/v1/admin/pipeline/revoke",
+        json={"video_id": video_id},
+        timeout=10,
+    )
+    return _json({
+        "video_id": video_id,
+        "status_code": resp.status_code,
+        "response": resp.json() if resp.status_code < 400 else resp.text,
+    })
+
+
+@mcp.tool()
+def rerun_stage(video_id: str, stage: int) -> str:
+    """Re-run a specific pipeline stage (2-6) for a video."""
+    if stage < 2 or stage > 6:
+        return "Error: stage must be 2-6"
+
+    import httpx
+
+    resp = httpx.post(
+        f"{API_URL}/api/v1/admin/pipeline/rerun-stage",
+        json={"video_id": video_id, "stage": stage},
+        timeout=10,
+    )
+    return _json({
+        "video_id": video_id,
+        "stage": stage,
+        "status_code": resp.status_code,
+        "response": resp.json() if resp.status_code < 400 else resp.text,
+    })
+
+
+@mcp.tool()
+def queue_depth() -> str:
+    """Check Celery queue depth and active/reserved/scheduled tasks."""
+    r = _redis()
+    queue_len = r.llen("celery")
+
+    # Inspect via celery CLI inside the worker container
+    client = _docker()
+    try:
+        worker = client.containers.get("chrysopedia-worker")
+        active_result = worker.exec_run(
+            "celery -A worker inspect active --json --timeout=5"
+        )
+        active = active_result.output.decode() if active_result.exit_code == 0 else "error"
+
+        reserved_result = worker.exec_run(
+            "celery -A worker inspect reserved --json --timeout=5"
+        )
+        reserved = reserved_result.output.decode() if reserved_result.exit_code == 0 else "error"
+    except Exception as e:
+        active = f"error: {e}"
+        reserved = f"error: {e}"
+
+    return _json({
+        "celery_queue_length": queue_len,
+        "active_tasks": active,
+        "reserved_tasks": reserved,
+    })
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 2: Infrastructure Health
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def worker_health() -> str:
+    """Check Celery worker status: online, pool size, active task."""
+    client = _docker()
+    try:
+        worker = client.containers.get("chrysopedia-worker")
+        result = worker.exec_run(
+            "celery -A worker inspect ping --json --timeout=5"
+        )
+        ping = result.output.decode() if result.exit_code == 0 else "no response"
+
+        stats_result = worker.exec_run(
+            "celery -A worker inspect stats --json --timeout=5"
+        )
+        stats = stats_result.output.decode() if stats_result.exit_code == 0 else "error"
+    except docker.errors.NotFound:
+        return _json({"status": "container not found", "container": "chrysopedia-worker"})
+    except Exception as e:
+        return _json({"status": "error", "error": str(e)})
+
+    return _json({"ping": ping, "stats": stats})
+
+
+@mcp.tool()
+def container_status() -> str:
+    """Show status of all Chrysopedia Docker containers."""
+    client = _docker()
+    containers = client.containers.list(all=True, filters={"name": "chrysopedia"})
+
+    result = []
+    for c in containers:
+        health = "unknown"
+        if c.attrs.get("State", {}).get("Health"):
+            health = c.attrs["State"]["Health"].get("Status", "unknown")
+
+        result.append({
+            "name": c.name,
+            "status": c.status,
+            "health": health,
+            "image": c.image.tags[0] if c.image.tags else "untagged",
+            "started_at": c.attrs.get("State", {}).get("StartedAt", ""),
+        })
+
+    return _json(sorted(result, key=lambda x: x["name"]))
+
+
+@mcp.tool()
+def service_health() -> str:
+    """Combined health check: DB, Redis, Qdrant, Ollama, API, Worker."""
+    checks = {}
+
+    # DB
+    try:
+        with _db() as conn:
+            with conn.cursor() as cur:
+                cur.execute("SELECT 1")
+        checks["postgresql"] = "healthy"
+    except Exception as e:
+        checks["postgresql"] = f"error: {e}"
+
+    # Redis
+    try:
+        r = _redis()
+        r.ping()
+        checks["redis"] = "healthy"
+    except Exception as e:
+        checks["redis"] = f"error: {e}"
+
+    # Qdrant
+    try:
+        import httpx
+
+        resp = httpx.get(f"{QDRANT_URL}/healthz", timeout=5)
+        checks["qdrant"] = "healthy" if resp.status_code == 200 else f"status {resp.status_code}"
+    except Exception as e:
+        checks["qdrant"] = f"error: {e}"
+
+    # API
+    try:
+        import httpx
+
+        resp = httpx.get(f"{API_URL}/health", timeout=5)
+        checks["api"] = "healthy" if resp.status_code == 200 else f"status {resp.status_code}"
+    except Exception as e:
+        checks["api"] = f"error: {e}"
+
+    # Worker (via container health)
+    try:
+        client = _docker()
+        worker = client.containers.get("chrysopedia-worker")
+        health = worker.attrs.get("State", {}).get("Health", {}).get("Status", "unknown")
+        checks["worker"] = health
+    except Exception as e:
+        checks["worker"] = f"error: {e}"
+
+    all_healthy = all(v == "healthy" for v in checks.values())
+    return _json({"overall": "healthy" if all_healthy else "degraded", "services": checks})
+
+
+@mcp.tool()
+def restart_service(service: str) -> str:
+    """Restart a Chrysopedia Docker container by short name.
+    Valid: api, worker, watcher, web, db, redis, qdrant, ollama"""
+    name_map = {
+        "api": "chrysopedia-api",
+        "worker": "chrysopedia-worker",
+        "watcher": "chrysopedia-watcher",
+        "web": "chrysopedia-web-8096",
+        "db": "chrysopedia-db",
+        "redis": "chrysopedia-redis",
+        "qdrant": "chrysopedia-qdrant",
+        "ollama": "chrysopedia-ollama",
+    }
+    container_name = name_map.get(service)
+    if not container_name:
+        return f"Unknown service '{service}'. Valid: {', '.join(name_map.keys())}"
+
+    try:
+        client = _docker()
+        container = client.containers.get(container_name)
+        container.restart(timeout=30)
+        return _json({"restarted": container_name, "status": "ok"})
+    except Exception as e:
+        return _json({"error": str(e), "container": container_name})
+
+
+@mcp.tool()
+def tail_logs(service: str = "worker", lines: int = 50) -> str:
+    """Get recent log lines from a Chrysopedia container.
+    service: api, worker, watcher, web. Default: worker."""
+    name_map = {
+        "api": "chrysopedia-api",
+        "worker": "chrysopedia-worker",
+        "watcher": "chrysopedia-watcher",
+        "web": "chrysopedia-web-8096",
+    }
+    container_name = name_map.get(service)
+    if not container_name:
+        return f"Unknown service '{service}'. Valid: {', '.join(name_map.keys())}"
+
+    try:
+        client = _docker()
+        container = client.containers.get(container_name)
+        logs = container.logs(tail=lines, timestamps=True).decode("utf-8", errors="replace")
+        # Truncate to 40KB to stay within MCP limits
+        if len(logs) > 40000:
+            logs = logs[-40000:]
+        return logs
+    except Exception as e:
+        return f"Error: {e}"
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 3: Content & Data
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def video_stats() -> str:
+    """Counts of technique pages, source videos, creators, with per-creator breakdown."""
+    totals = _query("""
+        SELECT
+            (SELECT count(*) FROM technique_pages) as technique_pages,
+            (SELECT count(*) FROM source_videos) as source_videos,
+            (SELECT count(*) FROM creators) as creators,
+            (SELECT count(*) FROM key_moments) as key_moments
+    """)
+
+    by_creator = _query("""
+        SELECT c.name, c.slug,
+            count(DISTINCT tp.id) as pages,
+            count(DISTINCT sv.id) as videos
+        FROM creators c
+        LEFT JOIN technique_pages tp ON tp.creator_id = c.id
+        LEFT JOIN source_videos sv ON sv.creator_id = c.id
+        GROUP BY c.id, c.name, c.slug
+        ORDER BY pages DESC, videos DESC
+    """)
+
+    by_status = _query("""
+        SELECT processing_status::text as status, count(*) as count
+        FROM source_videos
+        GROUP BY processing_status
+        ORDER BY count DESC
+    """)
+
+    return _json({
+        "totals": totals[0] if totals else {},
+        "by_creator": by_creator,
+        "by_processing_status": by_status,
+    })
+
+
+@mcp.tool()
+def list_videos(
+    status: str = "",
+    creator: str = "",
+    limit: int = 50,
+) -> str:
+    """List source videos, optionally filtered by processing status and/or creator slug."""
+    sql = """
+        SELECT sv.id, sv.filename, sv.processing_status::text as status,
+               c.name as creator, c.slug as creator_slug, sv.updated_at
+        FROM source_videos sv
+        LEFT JOIN creators c ON sv.creator_id = c.id
+    """
+    conditions = []
+    params: list = []
+
+    if status:
+        conditions.append("sv.processing_status = %s")
+        params.append(status)
+    if creator:
+        conditions.append("c.slug = %s")
+        params.append(creator)
+
+    if conditions:
+        sql += " WHERE " + " AND ".join(conditions)
+    sql += " ORDER BY sv.updated_at DESC LIMIT %s"
+    params.append(limit)
+
+    rows = _query(sql, tuple(params))
+    return _json({"count": len(rows), "videos": rows})
+
+
+@mcp.tool()
+def list_technique_pages(
+    creator: str = "",
+    format: str = "",
+    limit: int = 50,
+) -> str:
+    """List technique pages, optionally filtered by creator slug and/or format (v1/v2)."""
+    sql = """
+        SELECT tp.id, tp.title, tp.slug, tp.topic_category,
+               tp.body_sections_format, c.name as creator, c.slug as creator_slug,
+               tp.updated_at
+        FROM technique_pages tp
+        LEFT JOIN creators c ON tp.creator_id = c.id
+    """
+    conditions = []
+    params: list = []
+
+    if creator:
+        conditions.append("c.slug = %s")
+        params.append(creator)
+    if format:
+        conditions.append("tp.body_sections_format = %s")
+        params.append(format)
+
+    if conditions:
+        sql += " WHERE " + " AND ".join(conditions)
+    sql += " ORDER BY tp.updated_at DESC LIMIT %s"
+    params.append(limit)
+
+    rows = _query(sql, tuple(params))
+    return _json({"count": len(rows), "pages": rows})
+
+
+@mcp.tool()
+def list_creators() -> str:
+    """List all creators with video and technique page counts."""
+    rows = _query("""
+        SELECT c.id, c.name, c.slug, c.genre_tags,
+            count(DISTINCT sv.id) as video_count,
+            count(DISTINCT tp.id) as page_count
+        FROM creators c
+        LEFT JOIN source_videos sv ON sv.creator_id = c.id
+        LEFT JOIN technique_pages tp ON tp.creator_id = c.id
+        GROUP BY c.id, c.name, c.slug, c.genre_tags
+        ORDER BY page_count DESC, video_count DESC
+    """)
+    return _json({"count": len(rows), "creators": rows})
+
+
+@mcp.tool()
+def search_content(query: str, scope: str = "all", limit: int = 10) -> str:
+    """Run a search query against the Chrysopedia API. scope: all, topics, creators."""
+    import httpx
+
+    resp = httpx.get(
+        f"{API_URL}/api/v1/search",
+        params={"q": query, "scope": scope, "limit": limit},
+        timeout=15,
+    )
+    if resp.status_code != 200:
+        return f"Search failed: {resp.status_code} {resp.text}"
+    return _json(resp.json())
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 4: Pipeline Observability
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def recent_events(limit: int = 20, video_id: str = "", stage: str = "") -> str:
+    """Last N pipeline events. Optionally filter by video_id or stage name."""
+    sql = """
+        SELECT pe.id, pe.stage, pe.event_type, pe.model,
+               pe.prompt_tokens, pe.completion_tokens, pe.duration_ms,
+               pe.created_at, sv.filename
+        FROM pipeline_events pe
+        LEFT JOIN pipeline_runs pr ON pe.pipeline_run_id = pr.id
+        LEFT JOIN source_videos sv ON pr.source_video_id = sv.id
+    """
+    conditions = []
+    params: list = []
+
+    if video_id:
+        conditions.append("pr.source_video_id = %s")
+        params.append(video_id)
+    if stage:
+        conditions.append("pe.stage = %s")
+        params.append(stage)
+
+    if conditions:
+        sql += " WHERE " + " AND ".join(conditions)
+    sql += " ORDER BY pe.created_at DESC LIMIT %s"
+    params.append(limit)
+
+    rows = _query(sql, tuple(params))
+    return _json({"count": len(rows), "events": rows})
+
+
+@mcp.tool()
+def token_usage(video_id: str = "", stage: str = "") -> str:
+    """Token consumption summary. Optionally filter by video_id or stage."""
+    sql = """
+        SELECT pe.stage,
+               count(*) as calls,
+               coalesce(sum(pe.prompt_tokens), 0) as total_prompt_tokens,
+               coalesce(sum(pe.completion_tokens), 0) as total_completion_tokens,
+               coalesce(sum(pe.prompt_tokens), 0) + coalesce(sum(pe.completion_tokens), 0) as total_tokens
+        FROM pipeline_events pe
+        LEFT JOIN pipeline_runs pr ON pe.pipeline_run_id = pr.id
+        WHERE pe.event_type = 'llm_call'
+    """
+    params: list = []
+    if video_id:
+        sql += " AND pr.source_video_id = %s"
+        params.append(video_id)
+    if stage:
+        sql += " AND pe.stage = %s"
+        params.append(stage)
+
+    sql += " GROUP BY pe.stage ORDER BY total_tokens DESC"
+
+    rows = _query(sql, tuple(params))
+    grand_total = sum(r.get("total_tokens", 0) for r in rows)
+    return _json({"by_stage": rows, "grand_total_tokens": grand_total})
+
+
+@mcp.tool()
+def processing_times(limit: int = 10) -> str:
+    """Processing duration for recent pipeline runs (start to last event)."""
+    rows = _query("""
+        SELECT pr.id, sv.filename, c.name as creator,
+               min(pe.created_at) as started_at,
+               max(pe.created_at) as finished_at,
+               EXTRACT(EPOCH FROM (max(pe.created_at) - min(pe.created_at)))::int as duration_seconds,
+               count(pe.id) as event_count
+        FROM pipeline_runs pr
+        JOIN source_videos sv ON pr.source_video_id = sv.id
+        LEFT JOIN creators c ON sv.creator_id = c.id
+        JOIN pipeline_events pe ON pe.pipeline_run_id = pr.id
+        GROUP BY pr.id, sv.filename, c.name
+        ORDER BY max(pe.created_at) DESC
+        LIMIT %s
+    """, (limit,))
+    return _json({"count": len(rows), "runs": rows})
+
+
+@mcp.tool()
+def debug_mode_toggle(enable: bool | None = None) -> str:
+    """Get or set pipeline debug mode. Pass enable=true/false to change, omit to check."""
+    r = _redis()
+    key = "chrysopedia:debug_mode"
+
+    if enable is None:
+        current = r.get(key)
+        return _json({"debug_mode": current == "1" if current else False})
+
+    r.set(key, "1" if enable else "0")
+    return _json({"debug_mode": enable, "set": True})
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 5: Qdrant / Embeddings
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def qdrant_stats() -> str:
+    """Qdrant collection point counts and index health."""
+    import httpx
+
+    try:
+        resp = httpx.get(f"{QDRANT_URL}/collections", timeout=10)
+        collections_data = resp.json()
+        result = []
+
+        for col in collections_data.get("result", {}).get("collections", []):
+            name = col["name"]
+            detail_resp = httpx.get(f"{QDRANT_URL}/collections/{name}", timeout=10)
+            detail = detail_resp.json().get("result", {})
+            result.append({
+                "name": name,
+                "points_count": detail.get("points_count", 0),
+                "vectors_count": detail.get("vectors_count", 0),
+                "status": detail.get("status", "unknown"),
+                "segments_count": detail.get("segments_count", 0),
+            })
+
+        return _json({"collections": result})
+    except Exception as e:
+        return _json({"error": str(e)})
+
+
+@mcp.tool()
+def reindex_all() -> str:
+    """Trigger full re-embedding of all technique pages and key moments via the API."""
+    import httpx
+
+    resp = httpx.post(f"{API_URL}/api/v1/admin/pipeline/reindex-all", timeout=30)
+    return _json({
+        "status_code": resp.status_code,
+        "response": resp.json() if resp.status_code < 400 else resp.text,
+    })
+
+
+@mcp.tool()
+def reindex_video(video_id: str) -> str:
+    """Re-embed content for a specific video (re-runs stage 6)."""
+    return rerun_stage(video_id=video_id, stage=6)
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# DOMAIN 6: Prompt Management
+# ═══════════════════════════════════════════════════════════════════════════════
+
+
+@mcp.tool()
+def list_prompts() -> str:
+    """List current prompt files with sizes and SHA-256 hashes."""
+    prompts_dir = PROMPTS_PATH
+    if not os.path.isdir(prompts_dir):
+        return _json({"error": f"Prompts directory not found: {prompts_dir}"})
+
+    result = []
+    for fname in sorted(os.listdir(prompts_dir)):
+        fpath = os.path.join(prompts_dir, fname)
+        if not os.path.isfile(fpath):
+            continue
+        with open(fpath, "rb") as f:
+            content = f.read()
+        result.append({
+            "filename": fname,
+            "size_bytes": len(content),
+            "sha256": hashlib.sha256(content).hexdigest()[:16],
+            "lines": content.count(b"\n"),
+        })
+
+    return _json({"count": len(result), "prompts": result})
+
+
+@mcp.tool()
+def prompt_diff(filename: str) -> str:
+    """Compare a current prompt file to its most recent .bak version.
+    Returns a summary of changes (not full diff to save tokens)."""
+    prompts_dir = PROMPTS_PATH
+    current_path = os.path.join(prompts_dir, filename)
+
+    if not os.path.isfile(current_path):
+        return f"Prompt file not found: {filename}"
+
+    # Find most recent .bak
+    bak_files = sorted(
+        [f for f in os.listdir(prompts_dir) if f.startswith(filename) and f.endswith(".bak")],
+        reverse=True,
+    )
+
+    if not bak_files:
+        return _json({"filename": filename, "backup_found": False, "message": "No .bak file found"})
+
+    bak_path = os.path.join(prompts_dir, bak_files[0])
+
+    with open(current_path) as f:
+        current = f.readlines()
+    with open(bak_path) as f:
+        backup = f.readlines()
+
+    import difflib
+
+    diff = list(difflib.unified_diff(backup, current, fromfile=bak_files[0], tofile=filename, n=1))
+
+    added = sum(1 for l in diff if l.startswith("+") and not l.startswith("+++"))
+    removed = sum(1 for l in diff if l.startswith("-") and not l.startswith("---"))
+
+    # Truncate diff output
+    diff_text = "".join(diff[:200])
+    if len(diff) > 200:
+        diff_text += f"\n... ({len(diff) - 200} more lines)"
+
+    return _json({
+        "filename": filename,
+        "backup": bak_files[0],
+        "current_lines": len(current),
+        "backup_lines": len(backup),
+        "lines_added": added,
+        "lines_removed": removed,
+        "diff": diff_text,
+    })
+
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# Entry point
+# ═══════════════════════════════════════════════════════════════════════════════
+
+if __name__ == "__main__":
+    mcp.run(transport="streamable-http")