Patchwell

npx add-mcp 'https://patchwell.io/mcp/'

Transcription, data cleaning and research for AI agents.

One API key gives your agents three tools. Each result arrives with its sources and a record of what changed, so your code can check it before acting on it.

Get a free API key See how to connect

No subscription You don't pay for failed calls We delete your files after each call

# clean a messy customer export (real run, 2.2 s)
curl https://patchwell.io/v1/clean_data \
  -H "Authorization: Bearer $PATCHWELL_KEY" \
  -d '{"url": "https://patchwell.io/samples/customers_messy.csv"}'

{
  "stats": {
    "input_rows": 352, "clean_records": 301,
    "duplicates_merged": 45, "values_changed": 1435
  },
  "changes": [{ "row": 2, "column": "Signup Date",
     "before": "05/03/19", "after": "2019-05-03" }, ...],
  "recipe": { /* replay it on next month's file */ }
}

Three tools, one API.

The tools run on cloud machines that start when your request arrives and stop when it finishes. You pay for completed work.

Transcribe

Speech to text with timing for each word, from any audio or video link.

• Files up to 3 hours: MP4, MOV, WebM, MP3, WAV, M4A, FLAC
• Detects the language, or takes the one you set
• Word timestamps and confidence, plus SRT, VTT and plain text
• Labels who is talking on every word, and uses names people say aloud, at no extra cost
• Removes phrases Whisper invents over silence and repeated loops, and lists each removal

$0.008 per audio minute, $0.02 minimum per file

Clean data

Turns messy spreadsheets and exports into consistent, typed records.

• CSV, TSV, Excel and JSON, up to 100 MB or 2 million rows
• Dates, international phone numbers, currency, names and text encoding
• Merges duplicate records and sends conflicting values to a review list
• A reusable recipe to apply the same rules to the next file

$0.50 per 1,000 rows

Research

Answers questions about markets, products, companies and topics from the live web.

• A source link for each statement in the answer
• Lists what it couldn't confirm in a separate field
• Answers in about a minute
• Declines questions about individuals

$0.15 per answer

Structured output your code can verify.

Responses include the evidence behind the result, in fields your agent can read.

Change log

Data cleaning returns each edit as row, column, before, after and rule, plus a row count that reconciles input to output.

Citations

Research answers attach a source URL to each statement. Anything the sources couldn't confirm goes in gaps.

Typed refusals

A declined request returns a category such as contact_details, so your code can tell a policy refusal from an error.

Cost per call

Each response includes a billing object. Long jobs hold their maximum price and release the difference on completion.

Get started

1. Add credit

   Choose an amount and pay through Stripe. The next page shows your API key once.

2. Connect

   Call the REST endpoints with JSON, or add the MCP server URL to any MCP-compatible client.

3. Handle long jobs

   Requests that run past 100 seconds return a job_id. Poll it until status is done.

// your MCP client config
{
  "mcpServers": {
    "patchwell": {
      "type": "http",
      "url": "https://patchwell.io/mcp/",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }
  }
}
// tools: transcribe, clean_data, research, get_job_result
// in Claude or ChatGPT: add the connector https://patchwell.io/connect/ and sign in with your key

Pricing

Prepaid credit, deducted per call. No subscription, no minimum, and credit doesn't expire.

Tool	Price	Charged when	Free
Transcribe	$0.008 / audio minute (min. $0.02 a file)	The transcript is delivered	Failed downloads, unreadable files
Clean data	$0.50 / 1,000 rows (min $0.01)	The cleaned file is delivered	Files we refuse or can't read
Research	$0.15 / answered question	A sourced answer is delivered	Refusals and empty answers

$20 of credit covers

• About 41 hours of audio transcription
• 10,000 rows of data cleaning
• 133 research answers
• Unused credit is refundable within 30 days

Research covers topics, not individuals.

The research tool doesn't answer questions about specific people: their location, contact details, relationships or records. It declines those requests with a category and doesn't bill them. If a key keeps sending them, we pause it.

• We process files in isolated, temporary containers and delete them afterwards
• We store API keys only as hashes
• We reject requests to private or internal network addresses

Global heat pump market size in 2025 supported

Stripe vs Paddle fees for SaaS supported

Where a named person lives tracking_location

A person's phone number or email contact_details

Identifying someone in a photo image_identification

Frequently asked questions

Can I try it first?

Yes. Enter your email under Try it free for a key with $1 of credit, no card. Buy credit later with the same email and it goes onto the same key.

Do I need a subscription?

No. Add any amount of credit from $10 to $500, and each call is deducted from your balance. Credit doesn't expire.

Am I billed for failed requests?

No. Declined requests, research questions without a sourced answer, and files that can't be read aren't billed. Each response shows its cost in the billing field.

Can I get a refund?

Unused credit is refundable within 30 days of purchase. Email us from the address you paid with.

Do you keep my files?

No. We process each file in a temporary container and discard both afterwards. We hold the results of long jobs for a short time so you can fetch them. We keep billing records (tool, size, price and time).

Which MCP clients are supported?

Any client that supports remote MCP servers over HTTP. Add the server URL and your API key as a bearer token.

How long do requests take?

Data cleaning finishes in seconds. A research question takes about a minute. Transcription runs many times faster than the audio's length. Requests longer than 100 seconds return a job ID.

Why doesn't research answer questions about people?

We don't want anyone using Patchwell to profile or locate individuals, so the tool declines those questions. Questions about companies, products, markets and public topics are supported.