Skip to content
Maskflare
Talk to usStart free trial
Menu

For your agents · Maskflare Agent & MCP Security

Your Agents Read Everything. Make Sure the Model Doesn't.

An agent with a CRM tool, a mail tool and a file tool pulls customer records into its context without anyone pasting anything. Maskflare masks personal data and secrets in MCP tool calls and results, in both directions, so the model works with tokens and your systems keep the real values.
Start free trial Talk to us
28-day free trial or 10,000 requests, whichever comes first. No card needed.

What the CRM tool returns

{ "name": "Ada Lovelace",

"email": "ada@example.com",

"iban": "CH93 0076 2011 6238 5295 7",

"plan": "Business" }

Maskflare

What the model sees

{ "name": "mf_mask_170b",

"email": "mf_mask_5e1d",

"iban": "mf_mask_d7e0",

"plan": "Business" }

MCP server: crm · tools/call result · 3 values masked

574

Built-in detectors for personal data and secrets

47

Country packs, checksum-validated where the format has one

9

AI providers behind one gateway, plus the desktop agent and proxy

Sound familiar?

Why the usual answers fall short

“We reviewed the MCP servers.”

You reviewed what they can do. Not what data they return on Tuesday afternoon.

“Our DLP will catch it.”

If it blocks a tool result, the agent fails halfway through the task. Masking lets it finish.

“It's only internal tools.”

The model calling them isn't internal. Every tool result goes to the provider in the next request.

With chat, a person decides what to paste. An agent decides for itself, dozens of times per task, and nobody reviews each step.

Every MCP server you connect is a new door. A tool can return a full customer row, an invoice PDF or a .env file that the model was never meant to see, and that output becomes part of the next prompt.

How Maskflare Agent & MCP Security works

Route an MCP server through the Maskflare gateway at /v1/mcp/{server}. Tool-call arguments are masked before the server is called, and results are masked before the model sees them, for JSON and streamed responses alike. For agents you can't reconfigure, list the MCP hosts in the forward proxy and register the servers in the console. Tokens are stable, so the agent can reason about the same customer across steps, and real values come back in the final output.

  1. 01

    Register the MCP server

    In the console under Connections, MCP servers. Give it a name such as crm.

  2. 02

    Point the client at the gateway

    Use https://<your-gateway>/v1/mcp/crm instead of the server's own URL.

  3. 03

    Set the rules

    The same detectors, Exact Data Match lists and actions you use everywhere else. Block rules return a clean JSON-RPC error.

Benefits

Both directions

Arguments going to the tool and results coming back are both masked.

Agents keep working

Masking instead of blocking means the task completes; the model just never sees the raw values.

Stable tokens

The same customer always gets the same token, so multi-step reasoning still works.

Clean failures

A block rule returns a JSON-RPC error instead of a hang, so the agent can recover or ask for help.

Coding agents too

Hooks for Claude Code, Cursor and Codex CLI stop secrets at the prompt and mask tool input and output.

Admin assistant

The Maskflare admin MCP server answers questions like "which devices are unprotected?". Read-only by default.

Capabilities

Gateway route

/v1/mcp/{server} masks tools/call arguments before the server is contacted and masks results in JSON and SSE.

Proxy mode

List inspected MCP hosts in the forward proxy for clients that can't change their server URL.

Admin MCP server

uvx maskflare-mcp gives Claude Desktop, Claude Code or Cursor read tools for devices, usage, rules, the audit log and a masking preview. Write tools are opt-in and turn changes into proposals a person approves.

Set it up

MCP client configuration
{
  "mcpServers": {
    "crm": {
      "url": "https://<your-gateway>/v1/mcp/crm",
      "headers": { "Authorization": "Bearer mf_live_..." }
    }
  }
}
Admin MCP server
uvx maskflare-mcp
# read-only by default; set MASKFLARE_MCP_WRITE_TOOLS to allow proposals

What it does not do

We'd rather you hear it from us.

Explore related capabilities

See how Maskflare applies the same masking rules to related AI and data-protection workflows.

Maskflare Detection EngineMaskflare AI GatewayMaskflare Forward ProxyMaskflare Exact Data MatchMaskflare File SecurityMaskflare AI App Discovery & ControlMaskflare User CoachingMaskflare WorkforceMaskflare Browser ProtectionMaskflare Coding AgentsMaskflare SDKs & APIMaskflare Admin Console

Related guides and explainers

Frequently asked questions

Will masking confuse the agent?

No. Tokens are stable: the same customer always gets the same token, so the agent can reason about "customer mf_mask_170b" across steps, and Maskflare restores the value in the final output.

Which MCP clients work?

Any client that can point at an HTTP MCP endpoint can use the gateway route. Proxy mode covers clients that can't be reconfigured.

Does Maskflare store tool results?

No. Like prompts, tool results pass through and aren't logged. Only metadata and counts are recorded.

Can a rule stop a tool call entirely?

Yes. A block rule stops the call and returns a JSON-RPC error with the matching rule keys, never the values.

See it on your own data

Try Maskflare free.
Bring your hardest prompts.

28-day free trial or 10,000 requests, whichever comes first. Point one app at the gateway or enroll a few laptops,
and watch what gets masked. Or talk to us first.

Start free trial Talk to us