Skip to content

Adopt Claude org-wide. Your data stays yours.

DataFed is a BYOK gateway that detects sensitive data in every request and offers a one-click pseudonymized version — instead of blocking.

Start my free gatewayTalk to security

Email or SSO. No credit card. Your keys never leave your cell.

Illustrative — fixture data.

Blocking didn't stop the pasting. It hid it.

Most sensitive data reaches AI tools through personal accounts and copy-paste — exactly the traffic network blocking can't see. When the safe path is slower than the unsafe one, people rationally route around it. DataFed makes the safe path the fast path.

How a request moves through DataFed

One base-URL swap. Six things happen in the milliseconds after.

  1. 01

    Arrive

    Your app calls the gateway instead of the provider. One line changes:

    base_url = "https://cell-eu1.datafed.dev"

  2. 02

    Detect

    Deterministic detectors plus a small classifier find sensitive spans and their data classes.

  3. 03

    Decide

    Your admins' Cedar policies decide per class: allow, pseudonymize, or ask the user.

  4. 04

    Assist

    If a decision needs a human, the safe version is already computed — accepting it is one click.

  5. 05

    Forward

    Your provider key is decrypted just-in-time, the request goes out with surrogates in place.

  6. 06

    Learn

    Overrides and accept rates feed the eval loop, so noisy rules get flagged, not tolerated.

Read the architecture →

Your keys. Your cell. Our job is to touch as little as possible.

Bring your own keys

Envelope-encrypted, decrypted just-in-time per request. We never store plaintext keys, and there's no token markup — you pay Anthropic directly.

A cell of your own

Single-tenant data planes. Your prompts never share a process with another org.

Metadata-only control plane

The shared plane sees counts and classes, never content.

An SLO for staying quiet

We publish a false-interrupt-rate SLO and the eval numbers behind it. Every interruption spends your users' trust; we budget it.

Trust center: SOC 2 roadmap, subprocessors, pentest summary →

Built on the warning-science literature, not against it

Security warnings stop working by the second exposure. We designed for that finding instead of pretending it isn't true.

  1. [1]

    Making the safe choice the visually primary, lowest-effort option raised safe outcomes by ~30 percentage points at Chrome scale. Our one-click pseudonymized alternative is that button.

    — Felt et al., CHI 2015

  2. [2]

    fMRI shows habituation to a fixed warning begins on the second exposure. Our coaching cards rotate presentation.

    — Anderson et al., CHI 2015

  3. [3]

    Detect-pseudonymize-detokenize preserved task utility and user trust in LLM chat. It's the loop we run, hardened for enterprise.

    — Zhou et al. (Rescriber), CHI 2025

  4. [4]

    Block-with-override with logged justifications is proven enterprise practice. We extend it: the override telemetry tunes your policies.

    — Microsoft Purview DLP docs

Every provider your teams already use

One gateway in front of Anthropic, OpenAI, Hugging Face, AWS Bedrock, Google Vertex, Mistral — and anything OpenAI-compatible.

  • Anthropic
  • OpenAI
  • Hugging Face
  • AWS Bedrock
  • Google Vertex
  • Mistral
  • Azure OpenAI
  • Cohere
# before
client = Anthropic()
# after — policy applies org-wide
client = Anthropic(
  base_url="https://cell-eu1.datafed.dev")

Same key, same models, same bill. Policy in the middle.

Make the safe path the fast path.

Point your SDK at a DataFed cell and watch the first pseudonymized request go through in under five minutes.

Start my free gatewayTalk to security

We're onboarding a small group of design partners.

Email or SSO. No credit card. Transparent pricing publishes at beta — your keys, your provider bill.