com.eleyed/redactRedact PII before
it reaches the model
Your agent is about to send a customer document to an LLM. It should strip the personal data out first. Almost nothing does, because it is a manual detour.
eleyed exposes redaction as a tool your agent can call. Detection runs on our infrastructure — never a third-party model — and nothing you send is kept.
No card required. Purchased credits never expire.
{
"mcpServers": {
"redact": {
"url": "https://eleyed.com/api/mcp",
"headers": {
"X-API-Key": "rdk_live_your_key_here"
}
}
}
}Or, in Claude Code:
claude mcp add --transport http \
eleyed https://eleyed.com/api/mcp \
--header "X-API-Key: rdk_live_..."Two tools
Same detection and the same retention model as the REST API.
redact_textcosts credits by lengthRemoves personal data from a string. Returns the redacted text and a count of what was removed, by type — never the removed values themselves.
text — required.
entities — restrict to specific data types. Omit for all.
specific_terms — exact strings to remove as well, matched case-insensitively. For names the detector cannot know about.
redaction_style — how removed content is replaced. Defaults to a type label, e.g. <PERSON>.
score_threshold — minimum detection confidence, 0–1.
list_entity_typesfreeLists the data types available for the entities argument, so an agent can discover what is detectable. Costs no credits — charging for a discovery call would be hostile.
Data handling
You are piping customer text
through someone else's service
So the only questions that matter are what we keep and who else sees it. Both answers are short.
Nothing you send is stored
Processed in memory, discarded when the call returns. No document store, nothing written to disk, no backups of submitted content.
No third-party AI
Detection runs on our own infrastructure. Your text never reaches OpenAI, Anthropic, Google or any other outside model.
Tool arguments are never logged
Metering keys off the request headers rather than the body, so the text does not pass through logging on its way anywhere.
Only counts are kept
Operation type, credits used, timestamp, and how many of each entity type were found. Never the matched text.
What it costs
Credits, bought outright. No subscription, no seat licences, no minimum term.
redact_text
1 credit per call + 1 per 1,000 characters
list_entity_types
Free
Re-running the same text
Free — identical content is never charged twice
Ten free credits on signup, no card. Packs from $5.00, and purchased credits never expire.
PDFs go over the REST API
PDF redaction — including scanned pages and photographed documents, with OCR built in — runs over the REST API, where a file transfer is a file transfer. Same key, same credits, same retention model as the tools above.
It is not exposed as an MCP tool because a file does not fit in a tool call. Tool arguments are generated by the model, so a base64-encoded PDF would have to be emitted token by token — a ten-page document runs well past a hundred thousand tokens, and tool results have the same problem in reverse.
Before you wire it in
Is the text I send stored anywhere?
No. It is processed in memory and discarded when the call returns. There is no document store, nothing is written to disk, and there are no backups of submitted content. We keep the transaction record — credits used, timestamp, and counts of what was detected by type — but never the text.
Does it go to another AI provider?
No. Detection runs on eleyed’s own infrastructure. Your text is never passed to OpenAI, Anthropic, Google or any other outside model. That is the point of putting a redaction step in front of your model rather than behind someone else’s.
Can I redact PDFs?
Yes — over the REST API, including scanned pages and photographed documents, with OCR built in. Same API key, same credits, same retention model. It is not exposed as an MCP tool because a file does not fit in a tool call: tool arguments are generated by the model, so a base64-encoded PDF would have to be emitted token by token, and a ten-page document runs well past a hundred thousand tokens. Tool results have the same problem in reverse.
What happens when I run out of credits?
The tool returns an MCP error naming the top-up page, rather than an HTTP status your agent cannot interpret. Nothing is charged for a failed call, and re-running an identical redaction is always free.
Which protocol revision does it speak?
The 2026-07-28 revision, stateless Streamable HTTP. Clients still on the 2025 protocol are served from the same endpoint, so an older client works without any change on your side.
Wire it in and see what it finds
Ten free credits on signup — enough to run text through and read the output before you decide anything. No card, and nothing you send is kept.
Get an API key →