CortexDSP MCP
# CortexDSP MCP
**An open-source [MCP](https://modelcontextprotocol.io) server for Amazon DSP (Delivery Service Partner) operations.** Ask Claude about your scorecard, DNR investigations, DCR/POD metrics and more — right from Claude Desktop or claude.ai.
The demo tools work offline with zero setup. Plug in a **CortexDSP API key** and the server also exposes your real operation's data (DNR cases, who-delivered, scorecard comparisons, driver quality…) — securely proxied from the hosted API.
> Built for AMZL DSPs who are tired of Amazon's slow, fragmented dashboards.
## Quick start (Claude Desktop)
Add this to your `claude_desktop_config.json`:
```json
{
"mcpServers": {
"cortexdsp": {
"command": "npx",
"args": ["-y", "cortexdsp-mcp"]
}
}
}
```
Restart Claude Desktop. You now have the free demo tools. To unlock your own data, add your key:
```json
{
"mcpServers": {
"cortexdsp": {
"command": "npx",
"args": ["-y", "cortexdsp-mcp"],
"env": { "CORTEXDSP_API_KEY": "your-key-here" }
}
}
}
```
Get a key at **[cortexdsp.com](https://cortexdsp.com)**.
## Tools
### Free (no key, no data)
| Tool | What it does |
|---|---|
| `dsp_thresholds` | Scorecard metrics and their Fantastic targets (DCR, POD, DSC, DNR, LoR…). |
| `explain_metric` | What a metric means and how it's computed (`dcr`, `dsc`, `dnr`, `pod`, `lor`, `ementor`). |
| `dnr_response_format` | How a DNR (Did Not Receive) investigation is answered — the completion→location→detail cascade. Educational; sends nothing. |
| `pnov_benchmark` | Typical weekly ranges for lost / potential-not-on-van packages. |
| `return_budget` | How many returns you can have and still hit a target DCR. |
### With a CortexDSP key (your real data)
`dnr_list`, `missing_pnov`, `who_delivered`, `scorecard_compare`, `driver_quality`, `mentor_status`, `mail_list`, and more — scoped to your tenant. These are proxied from `api.cortexdsp.com`; this repo contains **no** business logic or credentials.
## How it works
```
Claude ──stdio──> cortexdsp-mcp (this repo)
├─ demo tools → answered locally, no data
└─ your-data tools → proxied to api.cortexdsp.com with your key
```
The same package serves everyone: without a key you get the demo tools; with a valid key + active subscription, your real tools are registered by proxy. If a key is missing or invalid, you simply fall back to the demo tools.
## Privacy & security
- This repo never contains API tokens, credentials, or internal endpoints.
- Your key lives only in your local MCP config and is sent to `api.cortexdsp.com` over HTTPS.
- Demo tools are fully self-contained and cannot read any customer data.
## Develop
```bash
npm install
npm test # unit tests (node:test)
npm start # run the server over stdio
```
## License
MIT © CortexDSP
TDQS
Scored across 5 tools
Each tool has a fairly distinct role: thresholds, metric explanation, DNR response guidance, PNOV benchmarks, and a return-budget calculator. The main overlap is that explain_metric includes DNR while dnr_response_format focuses on DNR procedure, and return_budget and dsp_thresholds both involve DCR targets, but descriptions provide enough separation.
Names mix noun phrases (dsp_thresholds, pnov_benchmark, return_budget), a verb phrase (explain_metric), and a longer noun phrase (dnr_response_format). The pattern is readable but not consistent, making it harder to predict tool names.
Five tools is well-scoped for a reference and calculator server focused on DSP scorecard metrics. Each tool appears to earn its place without being thin or excessive.
The surface covers metric targets, explanations, DNR response format, PNOV benchmarks, and a return-budget calculator, which is solid lifecycle coverage for the apparent educational domain. Minor gaps exist, such as no general-purpose calculator for arbitrary metric/target tradeoffs beyond returns, but agents can work around this.