@hiddencontent/mcp

@hiddencontent/mcp

An MCP server so an agent can check a document before it reads it.

Text hidden inside a document — white-on-white, zero-size fonts, text behind an image, off-page runs, comments, tracked deletions — is invisible to a person and read verbatim by a model. This server sits in front of that.

Design and the reasoning behind every decision: MCP.md.

Configure

Get a key (one command, returns a live token):

curl -fsS -X POST https://api.hiddencontent.ai/v1/signup

Then add the server. The JSON below works in Claude Desktop, Claude Code and Cursor unchanged:

{
  "mcpServers": {
    "hiddencontent": {
      "command": "npx",
      "args": ["-y", "@hiddencontent/mcp"],
      "env": {
        "HCS_API_TOKEN": "your-token",
        "HCS_CAPABILITIES": "tool-use"
      }
    }
  }
}

HCS_CAPABILITIES is worth setting. It is what turns on the checks for concealed text that instructs a model to send mail or call a tool (unauthorized-action, exfiltration). Declaring more can only raise severity, never lower it. Comma-separated: tool-use, network-egress, code-execution.

Tools

check_document

Give it one of path (best), url, or text.

Prefer a path or URL. Whole classes of concealment are properties of the file format and cannot be seen in extracted text, so a text submission is a deliberately weaker check and the response says so.

Returns a recommended action:

action	meaning
proceed	nothing concealed worth reporting; read the document normally
use_clean_text	concealed content was found and removed — read safe_text, not the file
ask_human	a person should look before the agent goes further

ask_human covers three cases, and the second is the one worth knowing about: concealed text shaped like an instruction to a model; concealed content that was deliberately retained in the text because removing it would have been data loss (so there is no safe version to hand over); and a degraded result, where the service did not answer and the check did not really happen.

get_evidence

Returns the concealed words from a previous check, so a person can judge them.

They are not in the default response. They come only from this explicit call, by which point the agent's routing decision is already made — an injection arriving then cannot change what it decided to do. The text is quoted inside a fence whose delimiter carries a token minted at analysis time, so a payload cannot close the block and escape into instruction position.

That last part is structural. The framing around it — declaring the block as data before and after, presenting spans as records — reduces compliance substantially but is not a proof.

What this package deliberately does not contain

No engine code, no rule pack, and no detection patterns. It calls /v1/analyze over HTTPS and its only dependency is the MCP SDK.

An earlier build used our client library, which carries an offline fallback, and that fallback dragged the rule pack — including the literal instruction-shape regexes — into the published bundle. Publishing a detection signature turns evasion from research into reading.

The consequence for you: with no connectivity you get "could not check this", not a partial local check. That is deliberate. If we are offline, we say so.

Build

npm run build  --workspace @hiddencontent/mcp   # tsc
npm run bundle --workspace @hiddencontent/mcp   # single-file bundle + dist/hiddencontent.mcpb
npx vitest run packages/mcp/test/

Against the live service:

HCS_CREDS=~/Desktop/hcs-credentials.txt npm run smoke:mcp

The unit tests prove the shaping logic against a fake client. They cannot prove the action an agent receives is right for a real document analysed by the deployed engine — every bug worth finding in this package was found that way, not by a fixture. smoke:mcp is that check, made repeatable.

bundle produces a self-contained dist/bundle.mjs (no node_modules required) and packs it into dist/hiddencontent.mcpb for one-click installation from Claude Desktop's Extensions tab. Verified 2026-08-28 by extracting the bundle into an empty directory and speaking MCP to it over stdio.