ParseRail MCP
# parserail-mcp
An [MCP](https://modelcontextprotocol.io) server for **[ParseRail](https://parserail.thecompound.tech)**, gives Claude, Cursor, and any Model Context Protocol client native tools to parse documents, extract fields, redact PII, analyze contracts, fight chargebacks, and enrich companies.
## Setup
Get a key at **[parserail.thecompound.tech](https://parserail.thecompound.tech)**. Add the server to your MCP client config. ParseRail has no free tier. Buy credits up front through a $20 pack or a plan from $19/mo. ParseRail charges a call only when it succeeds.
**Claude Code:**
```bash
claude mcp add parserail -e PARSERAIL_API_KEY=ksk_live_… -- npx -y parserail-mcp
```
**Claude Desktop / Cursor** (`mcpServers` config):
```json
{
"mcpServers": {
"parserail": {
"command": "npx",
"args": ["-y", "parserail-mcp"],
"env": { "PARSERAIL_API_KEY": "ksk_live_…" }
}
}
}
```
## Tools
| Tool | What it does |
| --- | --- |
| `parserail_parse` | Document (invoice/EOB/COI, PDF/image/text) → structured JSON |
| `parserail_extract` | Pull named fields out of any text |
| `parserail_classify` | Label text against your taxonomy |
| `parserail_summarize` | Summary + key points + action items |
| `parserail_redact` | Strip PII/PHI from text |
| `parserail_sentiment` | Sentiment, aspects, and themes |
| `parserail_contract` | Contract → parties, terms, obligations, risk flags |
| `parserail_chargeback` | Dispute details → representment packet |
| `parserail_enrich` | Domain or email → company profile |
| `parserail_account` | Credit balance |
Documents can be passed as `fileUrl`, raw `text`, or `fileBase64` + `fileMimeType`.
## Pricing
You pay with pay-per-call credits. You do not need a subscription. ParseRail charges you only when a call succeeds. See [parserail.thecompound.tech/docs](https://parserail.thecompound.tech/docs).
MIT, Compound Labs
TDQS
Scored across 41 tools
Most tools have distinct target use cases, but there are several overlapping pairs: summarize/minutes, classify/categorize, parse/invoice/receipt, and reply/outreach/review_reply. Reading the descriptions disambiguates them, but an agent is likely to need to compare closely before selecting.
The parserail_ prefix and snake_case style are consistent, which helps. However, the suffix convention is mixed: bare verbs (parse, split, rewrite), bare nouns (invoice, receipt, memory), and noun-first compounds (po_match, review_reply, product_copy), so there is no predictable verb_noun pattern.
At 41 tools, this is far beyond the typical well-scoped 3-15 tool range. Each endpoint may earn its place in the underlying API, but the MCP surface is too heavy for an agent to navigate easily. It feels like a full API dump rather than a curated set.
For a broad all-purpose AI-processing server, coverage is extensive: documents, audio, images, memory, outreach, finance, and account status are all present. However, because the domain is sprawling and many tools are one-shot generation or dead-end workflows, there is no clear lifecycle for created artifacts. Minor gaps like translation or custom voice selection are noticeable but not severe.