CTM VoiceAI MCP

Custom development, not an official CallTrackingMetrics product. Built and maintained by Jason Smith. For support, questions, or feedback, email jason.smith@ctm.com.

An MCP server (and Claude/Codex/pi skill) that reviews a CallTrackingMetrics (CTM) account's real call transcripts against its live VoiceAI agent instructions and produces prioritized, paste-ready prompt updates plus a fully rewritten prompt, rendered as a self-contained HTML report.

Highlights:

• No API keys at all. Login is CTM OAuth2 (device flow), and the analysis is performed by your MCP assistant itself - there is no external LLM call and no model API key.

• Data + renderer only. The server authenticates with CTM, fetches the agents and the call transcripts, and renders the final report. Your assistant does the thinking.

• Read-only. It never changes a live agent. Recommendations are proposals for a human to review and apply.

• PII-scrubbed output. Before anything is written or rendered, the server redacts emails, phone numbers, and detected caller/person names (2-3 word proper names) in the analysis prose, while preserving Markdown structure and the customer's own agent prompt.

• Works with any MCP-capable client: Claude Desktop, Claude Code, Codex, pi, or a local LLM agent.

The flow

1. Authenticate with CTM via OAuth device flow (ctm_voiceai_auth_login).

2. Get the VoiceAI agents and their instructions (ctm_voiceai_get_voice_bots).

3. Get the call activities and transcriptions (ctm_voiceai_get_calls). The response includes a plan sized to a default target of 500 calls (set target_calls to override, 0 for all), and the assistant dispatches one parallel subagent per batch so large call volumes are analyzed concurrently instead of serially.

4. The subagents compare the transcripts against the instructions, call by call, and return compact per-call JSON.

5. Your assistant merges the batches and writes the recommendations, including a full rewrite.

6. Write the report (ctm_voiceai_write_report), which saves the files, renders the HTML report, and opens it in your browser.

Related MCP server: llm-assure-mcp

What the HTML report contains

1. Topic analysis - canonical caller topics ranked by volume, with a High/Medium/Low voice-AI suitability rating and example calls.

2. Call analysis - the per-call extraction behind those topics.

3. Current agent prompt - the live instructions being reviewed.

4. Recommended prompt updates - a Coverage Map plus prioritized, copy-paste-ready prompt snippets.

5. Suggested rewritten prompt - a complete, self-contained rewrite with a Copy button.

Requirements

• Node.js 20+

• Python 3.9+ (standard library only; used for HTML rendering)

• Any MCP-capable client

Quick start

1. Get the code

git clone https://github.com/CTMJSON/ctm-voiceai-mcp.git
cd ctm-voiceai-mcp
npm install

2. Register it with your MCP client

Claude Code (one command):

claude mcp add ctmVoiceAI --scope user -- node "$(pwd)/src/index.js"

Claude Desktop - add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "ctmVoiceAI": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/ctm-voiceai-mcp/src/index.js"]
    }
  }
}

Codex - add this to ~/.codex/config.toml:

[mcp_servers.ctmVoiceAI]
command = "node"
args = ["/ABSOLUTE/PATH/TO/ctm-voiceai-mcp/src/index.js"]

Other MCP clients - use the same stdio command: node /ABSOLUTE/PATH/TO/ctm-voiceai-mcp/src/index.js.

3. Run it

Ask your assistant:

Review the VoiceAI agent for account <account id>: analyze the calls, compare them to the current agent prompt, recommend updates, and open the report.

The assistant walks the six steps above. It logs in to CTM if needed (you enter a short code in the browser), pulls the agents and transcripts, does the analysis, and finally writes and opens the HTML report.

Tools

Authentication is per-user: the server ships with a shared, public CTM OAuth client id (not a secret), but each user signs in with their own CTM login and only sees accounts that login can access.

Configuration (optional)

The defaults work out of the box. To change them, create ~/.config/ctm-voiceai/config.env:

# Optional tuning
CTM_VOICEAI_OPEN_REPORT=1        # 0 disables auto-opening the HTML report
CTM_VOICEAI_OUT_DIR=/custom/runs/dir
PYTHON_BIN=python3

Precedence: process environment, then ~/.config/ctm-voiceai/config.env, then any file named by CTM_VOICEAI_ENV_FILE. Never commit config.env or the stored tokens.

Output

Report runs are written to ~/.local/share/ctm-voiceai/runs/<account>-<timestamp>/:

voiceai_topic_analysis.html     the full report (opens automatically)
recommended_prompt_updates.md   the recommendations as Markdown
suggested_prompt_rewrite.md     the full rewritten agent prompt
analysis_artifacts.json         the raw analysis passed to the renderer
voiceai_topic_analysis.csv      ranked topics
run.json / run.log              run metadata and renderer log

Troubleshooting

• "NO_AUTH" - run ctm_voiceai_auth_login to sign in.

• Login code expired - device codes expire after about 25 minutes; just run ctm_voiceai_auth_login again for a new one.

• Report did not open - set CTM_VOICEAI_OPEN_REPORT=1 (default) or open the voiceai_topic_analysis.html path returned by ctm_voiceai_write_report.

• Thin analysis - accounts with few transcribed calls produce thin results. Check the reported call count before drawing conclusions.

Scope

Read-only against CTM (calls and voice-bot configuration). It never changes a live agent. No data leaves your machine except the analysis you explicitly write into the report - there is no external LLM call.

Support and feedback

This is a custom development, not an official CallTrackingMetrics product. Questions, bugs, feature requests, and feedback are welcome: jason.smith@ctm.com.