Gmc MCP

Google Merchant Center MCP (gmc-mcp)

MCP server for Google Merchant Center, built on Merchant API v1.

126 tools, MIT-licensed, pip install-able. Drive your Google Shopping

feed, products, inventory, Reports, promotions, returns, and account

configuration in natural language from Claude Desktop, Claude Code, or any

MCP-compatible client.

### Where this fits in the GMC + MCP landscape

| Option | Type | API | License | Install | Tools |

|---|---|---|---|---|---|

| Adzviser / Catchr / Windsor.ai / Pipedream | Paid hosted SaaS | varies | proprietary | sign-up | varies |

| archpeng/GMC-mcp-server | Self-hosted | Content API v2.1 (deprecated 2026-08) | none declared | git clone | 34 |

| `gmc-mcp` (this package) | Self-hosted | Merchant API v1 (future-proof) | MIT | `pip install gmc-mcp` | 126 |

Pick this if you want: future-proof API, an actual OSS license, PyPI install,

a larger toolset, audit log + rollback + dry-run safety, and credentials that

never leave your machine.

Built on Merchant API v1 (replaces Content API for Shopping; v1beta was discontinued 2026-02-28). Safe by default: every write is recorded to an append-only audit log with before snapshots, dry-run is one toggle away, and destructive bulk operations require an explicit confirm parameter.

Why this exists

Google Merchant Center (GMC) powers Google Shopping listings, Free Listings, Shopping Ads, Buy on Google checkout, YouTube Shopping, and more. Managing a feed of thousands of products through the GMC web UI is tedious, and writing one-off Python scripts against the Merchant API REST endpoints is slow.

There are existing MCP services for GMC, but they fit different niches:

• Paid hosted SaaS (Adzviser, Catchr, Windsor.ai, Pipedream) — your data is proxied through their servers; monthly subscription.
• `archpeng/GMC-mcp-server` — earliest self-hosted Python option (Feb 2026), but built on the deprecated Content API v2.1 which Google retires in August 2026, has no license declared, and isn't on PyPI.

gmc-mcp was built to fill the gap: a self-hosted package on the new Merchant API v1, MIT-licensed, on PyPI, with audit/rollback/dry-run safety and 126 tools across 19 modules.

The package gives Claude (or any MCP client) 126 tools that map directly onto the Merchant API v1 surface, so you can do things like:

• *"Show me every disapproved product, grouped by issue code."*
• *"Bump US prices 5% on these 40 SKUs."*
• *"Which products had impressions but zero clicks last week?"*
• *"Roll back audit ID 8fc8dc42… — that update broke something."*
• *"Compare this Shopify CSV against my GMC feed; what's missing?"*

What's exposed

126 tools across 19 modules + 4 meta tools.

CLI also exposes: gmc-mcp run, auth-init, register-gcp, webhook, describe, version.

Install

pip install gmc-mcp

For development:

git clone https://github.com/kiwoongeom/gmc-mcp

cd gmc-mcp pip install -e ".[dev,test]" pre-commit install

Quick start

1. Enable the Merchant API and create a service account

1. console.cloud.google.com → create / pick a project.
2. APIs & Services → Library → enable Merchant API.
3. IAM & Admin → Service Accounts → Create service account named gmc-mcp.
4. Open the new service account → Keys → Add key → Create new key → JSON. Save it somewhere safe.

2. Add the service account to your Merchant Center

1. merchants.google.com → top-right gear → Account access → People.
2. Add user → enter the gmc-mcp@PROJECT.iam.gserviceaccount.com email → grant Admin.
3. Note the 10-digit Merchant Center ID in the top-right.

3. Configure

cp .env.example .env

edit:

GMC_ACCOUNT_ID=1234567890

GMC_SERVICE_ACCOUNT_KEY=/abs/path/to/service-account.json

4. One-time GCP registration

Required before any v1 Merchant API call works for a new project.

gmc-mcp register-gcp --developer-email you@example.com

5. Smoke-test

gmc-mcp describe              # prints the resolved config

python examples/quickstart.py # makes one read against the Merchant API

6. Wire into Claude

Drop this into %APPDATA%/Claude/claude_desktop_config.json (Desktop) or ~/.claude/settings.json (Code):

{

"mcpServers": { "gmc": { "command": "gmc-mcp", "args": ["run"], "env": { "GMC_ACCOUNT_ID": "1234567890", "GMC_SERVICE_ACCOUNT_KEY": "/abs/path/to/service-account.json" } } } }

Restart Claude. Ask: *"What account am I connected to?"* — it should call gmc_describe_server.

OAuth instead of a service account

For acting on behalf of an end user (e.g. multi-tenant SaaS):

# 1. In GCP, create OAuth client credentials (type: Desktop), download client_secret.json.

gmc-mcp auth-init --client-secrets ./client_secret.json

follow the browser flow; a token JSON is saved to secrets/oauth-token.json.

2. Use it

export GMC_OAUTH_TOKEN=./secrets/oauth-token.json gmc-mcp run

Transports

gmc-mcp run                              # stdio (Claude Desktop default)

gmc-mcp run --transport sse --port 8000 # SSE / HTTP (remote, Docker, hosted)

Or via Docker:

docker build -t gmc-mcp .

docker run --rm -p 8000:8000 \ -e GMC_ACCOUNT_ID=1234567890 \ -v $PWD/secrets:/secrets \ -e GMC_SERVICE_ACCOUNT_KEY=/secrets/service-account.json \ gmc-mcp

Audit log + rollback

GMC_DRY_RUN=true blocks writes, but still sends read requests, including the read-only Reports API POST reports:search. Report tools return real data; Merchant API errors are surfaced to the caller rather than reported as zero rows.

Every write logs a JSONL entry to GMC_AUDIT_LOG (default: ./logs/audit.jsonl). Each entry contains:

• audit_id — UUID for lookup
• op, method, url, request_body, response_status, response_body
• before — pre-write snapshot when the tool fetched it (used by update_* and delete_*)
• dry_run — whether the request was actually sent

Inspect from inside Claude:

> Show me my last 5 GMC writes.

[gmc_audit_tail(limit=5) → ...]

Look up audit ID abc123.

[gmc_audit_lookup(audit_id="abc123") → ...]

Roll back audit ID abc123, that update was wrong.

[gmc_rollback(audit_id="abc123", confirm="ROLLBACK") → ...]

Dry-run

Two ways to enable:

GMC_DRY_RUN=true gmc-mcp run

> Set dry-run on and update the price of SKU1 to $19.99.

[gmc_set_dry_run(enabled=True) → ok] [gmc_update_regional_inventory(...) → {"_dry_run": true, ...}]

Reads always go through; only writes are skipped.

Reports query examples

Paste into gmc_query_report:

-- Top 50 by clicks last 7 days

SELECT offer_id, title, clicks, impressions, ctr FROM product_performance_view WHERE date BETWEEN '2026-04-23' AND '2026-04-30' ORDER BY clicks DESC LIMIT 50

-- Disapproved products with reason codes SELECT id, offer_id, title, item_issues FROM product_view WHERE aggregated_reporting_context_status = 'NOT_ELIGIBLE_OR_DISAPPROVED'

-- You're priced h