ads-mcp-server

Local MCP server exposing Google Ads + Meta Marketing performance data, campaign settings, and change history to Claude (Cowork) for live daily-dashboard workflows.

What it does

Three tools registered with MCP:

Architecture: pull ad×day raw once per platform per hour, cache to parquet, derive every aggregation in pandas. No API call per breakdown.

Prerequisites

• macOS / Linux
• Python 3.13 (via pyenv recommended)
• uv package manager: brew install uv
• Google Ads API access — see Google Ads API getting started
• Meta Marketing API access — see Marketing API getting started

Credential setup links

Install

cd ~/marketing-ds/ads-mcp-server
uv sync --extra dev

uv creates .venv/ and installs all deps pinned in pyproject.toml.

Configure credentials

Two options:

Option A — point at existing .env (recommended if you already have keys in ~/marketing-ds/decision_science/.env):

export ADS_MCP_ENV_FILE=/Users/jmacaggi/marketing-ds/decision_science/.env

Option B — local .env:

cp .env.example .env
# fill in the blanks

Required keys are listed in .env.example with comments explaining each.

Run

Locally for testing:

uv run ads-mcp-server

The server speaks MCP over stdio — Cowork (or any MCP client) spawns it on demand.

Connect to Claude (Cowork)

Add this block to ~/.claude/claude_desktop_config.json (create if missing):

{
  "mcpServers": {
    "ads": {
      "command": "uv",
      "args": [
        "--directory",
        "/Users/jmacaggi/marketing-ds/ads-mcp-server",
        "run",
        "ads-mcp-server"
      ],
      "env": {
        "ADS_MCP_ENV_FILE": "/Users/jmacaggi/marketing-ds/decision_science/.env"
      }
    }
  }
}

Restart Claude/Cowork. Tools get_google_ads_report, get_meta_ads_report, get_campaign_settings should appear.

No background daemon required — Cowork starts/stops the process per session.

Daily pre-warm (recommended for live dashboard)

Cache validity rule: a cache is fresh if it contains yesterday's date. Refreshes once per day. The first user query of the day triggers the refresh — and a 56-day pull on a large account can take 1-3 minutes.

To avoid that wait, schedule a pre-warm at 6am via macOS launchd:

# Install
cp /Users/jmacaggi/marketing-ds/ads-mcp-server/launchd/com.jmacaggi.adsmcp.prewarm.plist \
   ~/Library/LaunchAgents/

launchctl load ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist

# Verify it's scheduled
launchctl list | grep adsmcp

# Trigger immediately (test)
launchctl start com.jmacaggi.adsmcp.prewarm

# Logs
tail -f ~/marketing-ds/ads-mcp-server/logs/prewarm.stdout.log
tail -f ~/marketing-ds/ads-mcp-server/logs/$(date +%Y-%m-%d).log

What it does each morning at 6am:

1. Pulls Google Ads perf 56d, settings, change_event 29d → cache
2. (Meta currently disabled — see "Meta status" below)
3. Writes refresh stamp cache/google_lastrefresh.txt = today

By the time you open Cowork, Google data is hot. Tool calls return in <1s.

To uninstall:

launchctl unload ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist
rm ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist

Meta status (as of 2026-05-07)

Meta tools (get_meta_ads_report, get_campaign_settings(platform="meta"|"both")) are structurally complete but not yet verified end-to-end.

What works:

• Performance pull is chunked into 7-day windows (avoids the Service temporarily unavailable / subcode 1504044 "result too large" error)
• Ad-level data is split: level=campaign for the 56-day series, level=ad for yesterday-only (avoids 5-minute pagination)
• AdSet pull is filtered to effective_status IN [ACTIVE, PAUSED] (avoids paginating thousands of archived ad sets)

What blocked us:

• After the first big perf chunked pull, the AdSet pull hit Meta's hourly rate limit (code 17, subcode 2446079: "User request limit reached"). Cooldown is typically 10-60 minutes.

To re-test tomorrow morning (after quota resets):

# Remove --skip-meta from the launchd plist to enable Meta in pre-warm
sed -i '' '/<string>--skip-meta<\/string>/d' \
   ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist
launchctl unload ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist
launchctl load ~/Library/LaunchAgents/com.jmacaggi.adsmcp.prewarm.plist

# Or run manually
ADS_MCP_ENV_FILE=/Users/jmacaggi/marketing-ds/decision_science/.env \
   uv run python scripts/prewarm.py

If Meta keeps rate-limiting, fallback options (not yet implemented):

• Async report runs (async=True) for the perf pull
• Tighter effective_status filter (ACTIVE only, drop paused)

If you prefer a long-running background process (optional, not required for Cowork): use nohup uv run ads-mcp-server > /tmp/ads-mcp.log 2>&1 & or a launchd plist.

Optional: CSV override (skip the API)

Google Ads UI exports CSV reports without API quota limits. Drop a CSV into cache/external/ to override the API pull:

cache/external/google_2026-05-07.csv
cache/external/meta_2026-05-07.csv

If a matching CSV exists AND is newer than the parquet cache, the server loads it instead of calling the API. The response sets metadata.data_source = "csv_override" so Cowork knows.

CSV column schema must match the parquet (see src/ads_mcp_server/google_ads.py and meta_ads.py for column names: date, campaign_id, campaign_name, ad_id, ad_name, spend, impressions, clicks, conversions, ...).

Test

uv run pytest -v

All tests are mock-based — no network calls. Covers:

• classify_campaign Brand/Non-Brand/Other
• diagnose 5-state classifier
• 8-week same-DoW selector picks the right 8 dates
• WoW delta + zero-division
• actions[] filter for Meta
• Snapshot diff including null-old-value and pruning
• Missing creds returns clean error (no exception)

Logs

Every API call and error is logged to logs/YYYY-MM-DD.log (one file per day).

Troubleshooting

• google-ads install fails: ensure Python 3.13 (uv python pin 3.13) and pip install grpcio works on your system. On Apple Silicon: arch -arm64 uv sync.
• Meta token expired: regenerate the system user token in Business Settings; long-lived tokens last 60 days.
• Tool not appearing in Cowork: check ~/Library/Logs/Claude/mcp*.log for spawn errors. Confirm uv is in PATH for the GUI process (you may need a full path: which uv).
• Cache stale: delete cache/*.parquet to force fresh pull.

File map

src/ads_mcp_server/
  server.py        # MCP entry + tool handlers
  config.py        # env loading, validates creds
  google_ads.py    # 3 GAQL queries: perf, settings, change_event
  meta_ads.py      # Insights + AdSet pull
  snapshots.py     # Meta snapshot diff (Meta has no reliable change API)
  cache.py         # parquet + CSV override
  aggregate.py     # all pandas math
  classify.py      # Brand/NB/Other
  diagnose.py      # 5-state diagnosis
  date_ranges.py   # window resolution + 8wk DoW
  retry.py         # exponential backoff
  logging_setup.py # daily file logs
  schema.py        # response shape constants