Skip to content

Patchwell MCP

An agent handed a messy export usually does one of two things: refuses it, or quietly changes values and reports success. Patchwell is a narrow answer to that specific failure. Its cleaning tool returns every edit as a row, a column, a before, an after and the rule that made the change, plus a row count that reconciles the input with the output, and a recipe that replays the same rules on next month's file.

Server type: Remote (Streamable HTTP)
Endpoint:    https://patchwell.io/mcp/
Connector:   https://patchwell.io/connect/ (signed-in key for chat clients)
Auth:        Bearer API key (bought with prepaid credit)
Tools:       5 (transcribe, research, clean_data, get_job_result, listen_clip)
Pricing:     usage-based, per call, no subscription
Free path:   $1 of credit with an email, no card

What the endpoint says about itself

A keyless initialize on https://patchwell.io/mcp/ answers 200, not a 401. The server identifies itself as Patchwell 1.30.0 on protocol 2025-06-18, and tools/list returns five tools:

Tool What it does
clean_data Clean a messy spreadsheet or JSON export into typed, deduplicated records with a log of each change
research Answer a question from the live web, with each sentence citing the page it came from
transcribe Speech to text with word timestamps and confidence, speaker labels, SRT and VTT
get_job_result Fetch the result of a call that returned {"status": "running", "job_id": ...}, polled every 20 to 30 seconds
listen_clip Play a spot the speaker labels are unsure of. Free

The key arrives as a bearer token on the MCP connection, and the vendor also exposes REST endpoints on the same domain (/v1/clean_data, /v1/research, /v1/transcribe, /v1/jobs/{id}, /v1/credits/balance) for code that would rather call HTTP than speak MCP.

Clean data is the reason to keep the key

The published example is a 352-row customer export cleaned in 1.7 seconds: 301 clean records, 45 duplicate rows merged, 1,429 values changed, quality scored 0.30 before and 0.998 after, and each change returned as:

{"row": 2, "column": "Signup Date", "before": "05/03/19", "after": "2019-05-03"}

What the tool normalises, per the vendor: dates, phones, money, names, US street addresses, postal codes for 177 countries, units and broken encodings. Identifiers keep their leading zeros, and barcodes, ISBNs and IBANs are checked against their own check digits rather than guessed at. It merges duplicates and lists the likely ones it will not merge on a guess, which is the behaviour that makes the change log trustworthy: a refused merge is visible, not silent.

Input limits are stated: CSV, TSV, Excel and JSON up to 100 MB or two million rows. Output downloads as CSV or Excel, and the recipe is the part an operator reuses: apply the same decisions to next month's export instead of re-deciding them.

Research and transcribe, and what "cited" means here

research answers in about a minute and attaches a source URL to each statement, lists what it could not confirm in a separate gaps field, and reports a confidence level. It declines questions about individuals and returns a typed refusal category such as contact_details or tracking_location, so a policy refusal is distinguishable from an error in code. The sample answer on card processing fees returns the pricing, three source links, and a note that two claims were left out because the sources did not clearly support them.

transcribe takes any audio or video link up to three hours (MP4, MOV, WebM, MP3, WAV, M4A, FLAC), detects the language or takes the one you set, and returns words with start, end, probability and speaker labels, plus SRT, VTT and plain text. It removes the phrases Whisper invents over silence and repeated loops and lists every removal in flags, which is the same discipline as the change log applied to audio.

Pricing, read off the pricing table

Prepaid credit, deducted per call, no subscription and no expiry:

Tool Price Charged when
Transcribe $0.008 per audio minute, $0.02 minimum a file The transcript is delivered
Clean data $0.50 per 1,000 rows, $0.01 minimum The cleaned file is delivered
Research $0.15 per answered question A sourced answer is delivered

Credit is bought in whole dollars from $10 to $500, unused credit is refundable within 30 days, and failed downloads, refused files, refusals and empty answers are not charged. Each response carries a billing object, and a long job holds its maximum price and releases the difference on completion. $20 is presented as roughly 41 hours of audio, 10,000 rows cleaned, or 133 research answers. There is a $1 free key with no card.

Limits worth writing down

  • Trial credit is small. One dollar buys about 125 audio minutes or two research answers, enough to test the shape of the output and not enough to run a month.
  • 100 seconds is the line between a response and a job, so any real transcription run needs the polling loop, not a single call.
  • It deletes files after the call and holds only long-job results for a short window, so there is no archive to come back to; keep your own copy.
  • Research refuses people by policy. An operator expecting a contact lookup will get a typed refusal, by design.
  • No public repository and no self-host path, so the data path cannot be audited from outside beyond what the vendor states: isolated temporary containers, hashed keys, private and internal network addresses rejected.

Where it fits beside the other entries

The catalog already carries transcription utilities (Bulk Transcripts, transcript fetchers) and document automation. None of them do the cleaning step, and the cleaning step is where an operator's data actually breaks: the same customer twice under two spellings, a date column with three formats, a postal code that lost its leading zero. The pairing that earns this entry is the reconciliation: a change log for structured data, a source link for prose, and a flags list for audio, all three returned in fields an agent can check before it acts.

Frequently asked questions

What does Patchwell MCP cost?

Usage-based, per call: $0.008 an audio minute with a $0.02 minimum per file, $0.50 per 1,000 cleaned rows with a $0.01 minimum, and $0.15 per answered research question. Credit is bought from $10 to $500, does not expire, and is refundable for 30 days.

Is there a free way to test it?

Yes. An email address buys a key with $1 of credit and no card, which covers about 125 audio minutes or two research answers.

Does clean_data tell me what it changed?

Every edit comes back as a row, a column, a before value, an after value and the rule that made the change, plus stats that reconcile input rows to clean records and a recipe that replays the same decisions on a later file.

How do long jobs work?

Anything past 100 seconds returns a job_id, and get_job_result is polled every 20 to 30 seconds until the result is ready.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing