A stateless MCP server that turns any Google Sheet into a 4-entity CRM (Lead, Organization, Contact, Activity). Runs as a standalone HTTP service (FastMCP streamable-HTTP, default port 5002) or over stdio for Claude Desktop and other local MCP hosts — no gateway, no ngrok, no subprocess spawning required.
Any MCP host (Claude Desktop, your backend, an agent framework)
│ HTTP POST /mcp (JSON-RPC) — or stdio
▼
MCP-Excel :5002 (FastMCP streamable-HTTP)
│ gspread
▼
Google Sheets API
Every tool call includes an explicit sheet_id — there is no session state,
no "active sheet" variable. Multiple teams can use the same running instance
safely because nothing is shared between requests.
Four tabs, created on first connection if they don't exist. Existing tabs are used as-is; missing columns are appended without disturbing existing data.
| Tab | Primary Key | Key columns |
|---|---|---|
| Lead | Lead ID | Organization ID, Status, Segment, Opportunity Details, Last Contact Date |
| Organization | Organization ID | Organization Name, Industry Domain, Sub Domain, Company Type, Website |
| Contact | Contact ID | First Name, Last Name, Organization ID, Designation, Mobile, Email, LinkedIn Profile |
| Activity | Activity ID | Lead ID, Organization ID, Contact ID, Activity Type, Activity Sub-Type, Subject, Discussion Summary, Outcome |
IDs auto-increment continuing the existing pattern (LD001 → LD002,
ACT010 → ACT011). Falls back to UUID if no consistent pattern is found.
Each tab also gets Created At, Updated At, Deleted audit columns
appended if they aren't already present.
Headers are matched fuzzily (synonyms and near-duplicates are resolved), so
an existing sheet with slightly different column names still works. Run
tool_dedupe_headers (dry-run first) to merge genuine duplicate columns.
You need a creds.json service account key with Google Sheets API and
Google Drive API enabled.
Place it in this directory (the default path) or set GOOGLE_APPLICATION_CREDENTIALS
to its full path.
Never commit
creds.jsonto git. It is excluded in.gitignore; keep it local or mount it as a secret.
For each Google Sheet you want to connect: open the sheet → Share → add the service account email as Editor → Done.
The service account email is printed at startup and returned by get_sheet_info
on permission errors. It is also readable from creds.json:
python3 -c "import json; print(json.load(open('creds.json'))['client_email'])"pip install -r requirements.txtOptional — business-card OCR (Claude Vision) support:
pip install anthropic
export ANTHROPIC_API_KEY=...GOOGLE_APPLICATION_CREDENTIALS=creds.json python mcp_excel_server.pyVerify it's healthy:
curl -s -X POST http://localhost:5002/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' \
| python3 -m json.tooldocker build -t mcp-excel .
docker run -p 5002:5002 \
-v /path/to/creds.json:/secrets/creds.json \
-e GOOGLE_APPLICATION_CREDENTIALS=/secrets/creds.json \
mcp-excelSet MCP_TRANSPORT=stdio to use stdio instead of HTTP:
{
"mcpServers": {
"crm": {
"command": "python",
"args": ["/path/to/mcp-excel/mcp_excel_server.py"],
"env": {
"GOOGLE_APPLICATION_CREDENTIALS": "/path/to/mcp-excel/creds.json",
"MCP_TRANSPORT": "stdio"
}
}
}
}| Variable | Default | Notes |
|---|---|---|
GOOGLE_APPLICATION_CREDENTIALS |
creds.json |
Path to service account JSON |
MCP_TRANSPORT |
http |
http for server mode, stdio for Claude Desktop |
MCP_PORT |
5002 |
HTTP listen port |
ANTHROPIC_API_KEY |
— | Only for tool_extract_business_card |
All tools are stateless — every call requires sheet_id (full Google Sheets
URL or bare spreadsheet ID).
| Tool | Description |
|---|---|
tool_get_sheet_info |
Validate access, ensure all 4 tabs exist, confirm write access |
tool_add_record |
Add a record to Lead / Organization / Contact / Activity |
tool_list_records |
List records with optional column-based filters (paged) |
tool_search_records |
Keyword search across all columns |
tool_update_record |
Update fields on an existing record by primary-key ID |
tool_delete_record |
Soft-delete a record (sets Deleted to true) |
tool_get_related_records |
Look up records via foreign key (e.g. all Activities for a Lead) |
tool_get_pipeline_analytics |
Lead pipeline analytics: total, status breakdown, segment breakdown |
tool_dedupe_headers |
Find and merge duplicate columns (dry-run report first) |
tool_extract_business_card |
Extract contact fields from a business-card image via Claude Vision (needs ANTHROPIC_API_KEY) |
pip install pytest
pytest test_mcp_excel.py test_crm_local.py -vTests run fully offline (Google calls are mocked); no credentials needed.
| Problem | Fix |
|---|---|
status: failure, message: "" |
Sheet not found — share it with the service account as Editor |
Permission denied on sheet |
Re-share with the service account as Editor (not Viewer) |
Cannot connect to MCP-Excel service |
MCP-Excel not running — start it on :5002 |
sheet_id is required and was empty |
Always pass sheet_id — there is no implicit active sheet |
| Tab not found / wrong columns | get_sheet_info auto-creates missing tabs and appends missing columns |