Dialpad MCP Server

An MCP server that exposes Dialpad calls and transcripts as tools, for use from Claude Desktop and Claude Code.

Tools

Base API: https://dialpad.com/api/v2 (override with DIALPAD_BASE_URL for sandbox).

Related MCP server: Intercom MCP Server

Setup

cd ~/projects/dialpad-mcp
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

# Add your key (kept out of git via .gitignore)
cp .env.example .env
# then edit .env and set DIALPAD_API_KEY=...

The server auto-loads .env, so your API key never has to live in client config JSON.

Get a Dialpad API key

Dialpad Admin → Company Settings → Authentication → API keys (you need a Company Admin role). Paste the key into .env as DIALPAD_API_KEY.

Connect to Claude Code

From this project directory it's picked up automatically via .mcp.json. To register it for all projects (user scope):

claude mcp add dialpad -s user \
  -- /home/fredericktejada/projects/dialpad-mcp/.venv/bin/python -m dialpad_mcp

Then run claude mcp list to confirm it shows up, and /mcp inside a session.

Connect to Claude Desktop

Edit the config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add:

{
  "mcpServers": {
    "dialpad": {
      "command": "/home/fredericktejada/projects/dialpad-mcp/.venv/bin/python",
      "args": ["-m", "dialpad_mcp"],
      "cwd": "/home/fredericktejada/projects/dialpad-mcp"
    }
  }
}

On Windows + WSL, the desktop app can't call the Linux venv directly. Either run Claude Desktop's command through wsl.exe ("command": "wsl.exe", "args": ["-e", "/home/fredericktejada/projects/dialpad-mcp/.venv/bin/python", "-m", "dialpad_mcp"]) or run the server inside a native Windows Python install.

Restart Claude Desktop after editing.

Export transcripts to a Google Sheet

export_transcripts.py pulls call transcripts into a Google Sheet, one row per call.

# New sheet, last 7 days, recorded calls only
.venv/bin/python export_transcripts.py --days 7 --recorded-only

# Full rebuild of a tab (clears + rewrites)
.venv/bin/python export_transcripts.py --days 14 --recorded-only --sheet-id <ID> --tab Transcripts

# Append only NEW calls (dedup by Call ID; skips transcript fetch for known calls)
.venv/bin/python export_transcripts.py --append --days 2 --recorded-only --sheet-id <ID> --tab Transcripts

Daily scheduled append

run_daily.sh appends new calls to the live sheet (1Jlea4o_0g0h9HygILZV4XbN305E_fW9DWgD-qDh3q08). Installed in cron at 07:45 Manila:

45 7 * * * /home/fredericktejada/projects/dialpad-mcp/run_daily.sh >> .../cron.log 2>&1

--append reads existing Call IDs first and skips them, so the daily run only fetches transcripts for genuinely new calls (fast, idempotent — safe to overlap windows).

Columns: Call ID, Date started, Direction, Category, External/Internal number, Agent, Email, Talk (s), Total (s), Recorded, Moments, Transcript.

Google auth reuses the shared OAuth client at ~/.api-credentials/google/oauth-client.json. The account token lives in secrets/google-sheets-token.json (gitignored). To (re)connect a different Google account on WSL, use the manual paste flow:

.venv/bin/python g_auth.py --url      # open the printed URL, sign in, click Allow
#   copy the localhost redirect URL into code.txt
.venv/bin/python g_auth.py --finish   # saves the token

Key options: --days, --limit, --recorded-only, --target-id/--target-type, --office-id, --timezone (default Australia/Melbourne), --sheet-id, --title, --tab.

Analytics tabs (Overview / Call Reasons / Customer Type)

build_analytics.py adds three analysis tabs to the spreadsheet:

.venv/bin/python build_analytics.py --days 14 --limit 8000 \
    --sample-reasons 800 --sample-customers 800 \
    --sheet-id 1Jlea4o_0g0h9HygILZV4XbN305E_fW9DWgD-qDh3q08

• Overview — per agent/line: total calls, inbound/outbound, answered, avg talk, avg handle time, total talk hours. Built on the full call set (lines vs agents split by name).

• Call Reasons — recorded inbound transcripts classified (keyword rules + custom moments) into Booking / Voicemail / Theory / Certificate / Cancellation / Payment / Reschedule-transfer / Location / Group / Course-info / Other, with avg handle time per concern. Runs on a random sample (--sample-reasons), extrapolated to the population.

• AHT by Agent x Concern — matrix of each agent's average handle time per concern (cells show m:ss (n)), built from the same reason sample. Agents need ≥8 sampled calls.

• Customer Type — unique inbound numbers matched against aXcelerate (/contacts/search?q=<local#>) and HubSpot (crm/v3/objects/contacts/search, CONTAINS_TOKEN *<9 digits>* on phone+mobilephone). Returning = found in either. Runs on a random sample (--sample-customers).

Lookups are cached in .cache_lookups.json (gitignored) so re-runs are cheap. aXcelerate tokens come from ~/.api-credentials/.env; the HubSpot token from ~/projects/cprfa-tickets/secrets/.env (needs crm.objects.contacts.read scope). Volume note: ~8k call legs / 14 days, so reasons + customer type are sampled by design.

Daily CALLS + AHT tracker (build_daily_report.py)

Row-per-day tabs inspired by the ops "2026 REPORTS" sheet:

• Daily AHT — Day | AHT (all) | <each agent's AHT> (one row per day)

• Daily CALLS — Day | volume metrics (Admin/Sales inbound, outbound, missed, abandoned, voicemail) | Top Call Drivers (reason categories)

# Full 30-day (re)build — pulls in 5-day chunks (single 30d export won't generate)
.venv/bin/python build_daily_report.py --days 30
# Daily append of newly-completed days (run by run_daily.sh / cron)
.venv/bin/python build_daily_report.py --append

Dialpad can't generate a 30-day export in one shot, so fetch_chunked primes all day-range chunks at once (concurrent server-side generation) then collects each as it finishes (days_ago_start/days_ago_end). Today is excluded (partial); the daily cron adds each day once complete. Drivers reuse the transcript cache.

run_daily.sh (cron, 07:45 Manila) now appends transcripts + Daily AHT + Daily CALLS.

Notes

• list_calls uses the Stats export API, which is asynchronous — the tool blocks while polling (up to ~90s) then parses the returned CSV.

• Rate limits vary per endpoint (get_call is only ~10/min; transcripts ~1200/min).

• Secrets live in .env only — never commit it.