Patchwell MCP server
Commit Nov 12, 2026Enables AI agents to perform cited web research, transcribe audio/video with speaker labels, and clean spreadsheet data with change logs.
Also known as aidan-mccollum/patchwell
- Stars
- 0
- Tools
- 4
- Downloads, 30 days
- —
- Latest version
- 1.3.0
- Last commit
- Nov 12, 2026
- License
- MIT
What is Patchwell MCP?
Patchwell enables AI agents to perform cited web research, transcribe audio/video with speaker labels, and clean spreadsheet data with change logs. It exposes 4 tools, such as research, transcribe, clean_data.
Summary generated from the project's README
About Patchwell MCP server
Patchwell provides a suite of tools designed for AI agents to perform specialized tasks like cited web research, audio and video transcription, and spreadsheet data cleaning. It operates as a remote MCP server or via a plain HTTP API, allowing developers to integrate these capabilities into clients like Claude, ChatGPT, Cursor, and Windsurf. Users can connect by providing an API key, which can be obtained with a free trial credit.
The research tool generates answers where every sentence includes a citation and a confidence grade. For media, the transcribe tool handles audio or video URLs up to three hours long, providing word-level timestamps and speaker identification. The clean_data tool processes CSV, Excel, or JSON files, producing a detailed change log and a reusable recipe. This service uses a pay-per-call model rather than a subscription, making it suitable for targeted automation and specific developer workflows.
What you can do with it
- 1Perform cited web researchAsk an AI agent to research specific topics or market regulations with verifiable source links for every claim.
- 2Transcribe audio with speaker labelsConvert long interview recordings into text that identifies different speakers and provides word-level timestamps.
- 3Clean messy spreadsheet dataAutomate the process of fixing errors and removing duplicates in large CSV or Excel files while maintaining a log of all changes.
Written from the project's README. Every use case is checked against a quote from it.
How to install Patchwell MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall Patchwell MCP in Claude Code
$ claude mcp add --transport http patchwell https://patchwell.io/mcp/Install Patchwell MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"patchwell": {
"url": "https://patchwell.io/mcp/"
}
}
}Install Patchwell MCP in VS Code
.vscode/mcp.json
{
"servers": {
"patchwell": {
"type": "http",
"url": "https://patchwell.io/mcp/"
}
}
}Install Patchwell MCP in VS Code CLI
$ code --add-mcp '{"name":"patchwell","type":"http","url":"https://patchwell.io/mcp/"}'Install Patchwell MCP in Codex
$ codex mcp add patchwell --url https://patchwell.io/mcp/Install Patchwell MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.patchwell] url = "https://patchwell.io/mcp/"
Install Patchwell MCP in Gemini CLI
$ gemini mcp add -s user --transport http patchwell https://patchwell.io/mcp/Install Patchwell MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"patchwell": {
"url": "https://patchwell.io/mcp/",
"type": "http"
}
}
}Patchwell MCP tools 4
Detected from the READMEresearchtranscribeclean_dataget_job_result
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableaidan-mccollum/patchwell is public and reachablegithub.com/aidan-mccollum/patchwell ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-12github.com/aidan-mccollum/patchwell ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model4 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeinitialize answered with protocol 2025-06-18 in 534 msPass
README
View on GitHubTools
| Tool | What it returns | Price |
|---|---|---|
research |
An answer where each sentence cites the page it came from, what the sources couldn't confirm, and a confidence grade. Topics, markets, products, companies, regulations, published research. Declines questions about private individuals and about anyone's personal details; public figures' public roles and statements are fine. | $0.15 per answered question. Declined or unanswered: free |
transcribe |
Text, word-level start/end times and confidence, who said each word (Speaker 1, 2, ..., or a name when the recording says it in a set phrase like "My name is ...", with the quote as evidence), speaker turns, sentence segments, SRT and VTT with speaker tags. Audio or video URLs up to 3 hours. Stock phrases Whisper invents over silence ("Thanks for watching!") and repeated loops are removed, each with a flag holding the original text and timestamps. | $0.14 per audio hour (min. $0.02 a file) |
clean_data |
Clean rows from a CSV, Excel or JSON file, plus a log of every changed value, a review list, detected column types, data quality before and after, a JSON Schema and a reusable recipe. Exports to CSV and Excel. | $0.50 per 1,000 rows |
get_job_result |
The result of a long job (anything over about 100 seconds returns a job_id). |
Free |
From our own runs: a 44-minute interview took 85 seconds end to end, speaker labels included (one run, GPU start included). On 12 public recordings held out from our tuning (AMI, VoxConverse, Earnings-21 and SIMSAMU; 32 in all, 20 used for tuning), 97.1% of words went to the right speaker. That's a word-level measure, not DER, and words where two people overlap aren't scored. A messy 352-row customer file took about 2 seconds to clean, with 45 duplicates merged and 1,429 values fixed, each one logged. A research question typically takes 50 to 100 seconds (median about 70 seconds, up to about 3 minutes), plus 20 to 40 seconds when the model has to start after a quiet period.
Get a key
Free key with $1 of credit (more on launch day). It needs an email but no card. You get one per email and one per network every 30 days, and only a set number go out each day:
curl -X POST https://patchwell.io/v1/credits/trial \
-H "Content-Type: application/json" -d '{"email": "you@example.com"}'
Or buy credit at patchwell.io. Keys look like pw_live_....
Connect
Claude (web and desktop) or ChatGPT
Add a custom connector with the address https://patchwell.io/connect/. The app opens a Patchwell page where you paste
your key or get a free one, then sends you back. In Claude it's under Customize, then Connectors, then Add. In ChatGPT, use
developer mode.
Claude Code
Install the plugin, then run /mcp, pick patchwell and sign in. A Patchwell page asks for your key or gives you a
free one.
/plugin install patchwell --marketplace aidan-mccollum/patchwell
Or connect with a key you already have:
claude mcp add --transport http patchwell https://patchwell.io/mcp/ --header "Authorization: Bearer YOUR_KEY"
Gemini CLI
gemini extensions install https://github.com/aidan-mccollum/patchwell
Then run /mcp auth patchwell inside Gemini CLI to sign in.
Cursor, Windsurf and other clients that take JSON config
{
"mcpServers": {
"patchwell": {
"type": "http",
"url": "https://patchwell.io/mcp/",
"headers": { "Authorization": "Bearer YOUR_KEY" }
}
}
}
Claude Desktop (through the mcp-remote bridge)
{
"mcpServers": {
"patchwell": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://patchwell.io/connect"]
}
}
}
The first time Claude Desktop starts the bridge, a Patchwell page opens in your browser and asks for your key or gives you a free one. The key stays out of this file.
Plain HTTP
curl https://patchwell.io/v1/research \
-H "Authorization: Bearer $PATCHWELL_KEY" -H "Content-Type: application/json" \
-d '{"ask": "Which open-source licenses allow commercial use?"}'
What a research answer contains
| Field | What it holds |
|---|---|
answer |
The answer. Each sentence ends with the URL of the page it came from, in brackets |
claims |
Each sentence again, with its page (URL and title) and the quote found on that page, so your code can check it without fetching anything |
sources |
The pages the answer rests on, with their titles |
gaps |
What the sources couldn't confirm, and how many claims the check dropped |
confidence |
high, medium or low |
billing |
What the call cost: $0.15 for an answer, nothing when it declines or finds no sourced answer |
Before you see the answer, a second step checks each claim against the page it cites. The quote behind a claim has to appear on that page (string matching, no model involved), and then the same model judges whether the quote supports the claim. Claims that fail are dropped and counted in gaps. A claim that passes means the page says it, not that it's true, so open the cited page before you rely on anything important.
Also on Apify
The same tools run as Apify Actors, billed through your Apify account at Apify per-event prices, which can differ from the direct API: research · transcribe · clean data
Built for automations
-
Retries never charge twice: send an
Idempotency-Keyheader; a retry gets the first response back. -
Webhooks: add
callback_urland get the result POSTed when it's done, signed with HMAC-SHA256 and retried for about two hours. -
Presets: save a configuration once (
PUT /v1/presets/<name>) and send"preset": "<name>". -
Exports:
POST /v1/export?format=docx(orsrt,vtt,txt,md,csv) turns a transcript into a file. -
Speaker labels you can check: every word carries
spk_p_wrong, the chance its speaker label is wrong (checked on recordings held out from our tuning), andspeaker_reviewlists the spots worth a listen with links that open the file at that second. -
Stable results: every result has
schema_version; within v1, fields are only added. -
Talk-time analytics: each speaker's share, pace, turns, longest stretch, questions and interruptions.
-
Rename speakers in the Claude card, or with
POST /v1/renameandspeaker_nameson exports. -
Templates: n8n workflows (tested inside n8n) and a schedulable podcast-to-notes script at patchwell.io/automations.
How these numbers were measured, and where the tools fall short: patchwell.io/trust.
Privacy and limits
- Patchwell deletes input files (audio, video, spreadsheets) after each call. It keeps a long job's result up to 7 days for pickup, and the full response to a request sent with an
Idempotency-Keyfor 24 hours, so a retry gets the same answer. - Research questions aren't kept unless you turn that on: the box when you buy credit, the account setting (
POST /v1/account/privacy), or the headerX-Patchwell-Retain: on. Kept questions stay 30 days with contact details removed, and Patchwell uses them only to find where the tool goes wrong. - No model API company (OpenAI, Anthropic and the like) receives your questions: the model runs on GPUs Patchwell rents. Serper and Tavily receive the search queries the tool writes from your question, and Wikipedia's search and the GDELT news index receive its topic for lookups. Every provider is listed in the privacy policy.
- Research declines questions about private individuals and about anyone's personal details (where they live, contact details, family, health, finances, tracking someone). Public figures' public roles and statements are fine. If a key keeps sending declined questions, Patchwell pauses research on it for 24 hours.
- Speaker labels use voice embeddings computed in memory during one call and never stored. Send
"speakers": falseto skip them. - The tools refuse URLs that point at private or internal network addresses.
- If Patchwell shuts down, you get at least 30 days' notice and a refund of all unused credit you paid for. Good-faith testing, including trying to break the refusals, is welcome: see the terms.
About this repository
This repo holds the MCP Registry entry (server.json), the Claude Code plugin (plugins/patchwell), the Gemini CLI
extension (gemini-extension.json), the Cursor plugin (.cursor-plugin/) and these docs. The service runs at
patchwell.io.
Questions: patchwell.support@gmail.com
Patchwell MCP questions
How do I add Patchwell MCP to Claude Code?
Run claude mcp add --transport http patchwell https://patchwell.io/mcp/ in your terminal. Add --scope user to make it available in every project.
How do I add Patchwell MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
What tools does Patchwell MCP provide?
It exposes 4 tools, including research, transcribe, clean_data, get_job_result.
How much does the research tool cost?
It costs $0.15 per answered question, while declined or unanswered questions are free.
Does Patchwell store my input files?
Patchwell deletes input files such as audio, video, and spreadsheets after each call.
Can I use Patchwell for transcription?
Yes, it provides text, word-level timestamps, speaker identification, and exports to formats like SRT and VTT.
How does the data cleaning tool work?
It cleans rows from CSV, Excel, or JSON files and provides a log of every changed value and data quality metrics.