Open app
AI PRIVACY INFRASTRUCTURE LOCAL-FIRST

YOUR AI NEEDS
CONTEXT.
NOT IDENTITY.

SHREDDAR removes unnecessary sensitive information before data reaches cloud AI, then reconstructs only authorized references when the response returns.

SHRED SOMETHING ↗OPEN APP →
0.1 ALPHABrowser-local previewSelf-hosted gatewayNo raw telemetry
LIVE / BROWSER LOCAL POLICY ACTIVE
ORIGINALACTIONPROTECTED
Mira ChenCOMPILE<PERSON_01>PERSON
mira@northstar.exampleREMOVEREMOVEDEMAIL
€82,471COMPILE€80K–€90KFINANCE
demo_key_••••••BLOCK<API_SECRET_PRESENT>SECRET
OUTBOUND CONTEXT−47.3%IDENTITY RETAINED LOCALLY
PRIVATE WORLDSHREDDAR EDGEAI WORLD
INPUTPRIVATE DATA
→
PRIVACY COMPILER
→
OUTPUTMINIMUM USEFUL CONTEXT
01

THE EXPOSURE BOUNDARY

Cloud intelligence shouldn't require unnecessary disclosure.

Every model, agent and tool call creates a decision: what actually needs to leave your environment? SHREDDAR makes that decision inspectable and enforceable.

TRUSTED ENVIRONMENTContractsCustomer recordsSource codeWallet operations
RAWSHREDDARPOLICY / VAULT / ROUTEPROTECTED
EXTERNAL INTELLIGENCECloud modelsAI agentsMCP servicesExternal APIs

SEE EXACTLY WHAT YOUR AI SEES. Context crosses the line. Identity does not have to.

02

THE PIPELINE

Controlled destruction.
Useful intelligence.

01

DETECT

Locate identity, secrets, exact values and linkable facts.

02

SHRED

Remove information the requested task does not require.

03

ABSTRACT

Preserve meaning while reducing identifying specificity.

04

SHARD

Create independently routable semantic compartments.

05

INFER

Send only policy-approved protected representations.

06

STITCH

Reconstruct authorized references inside the local boundary.

03

LIVE PLAYGROUND

Inspect the exact context
allowed to leave.

The preview below runs in browser memory. Use synthetic or authorized content—never real credentials. Nothing is sent to our backend.

RUNS IN THIS TAB · ORIGINAL NEVER SENT
01 / ORIGINAL
02 / PROTECTED
Scanning locally…
0 ENTITIES−0% CONTEXT
03 / AI RESPONSE
Waiting for protected context…
MODEL SEES PLACEHOLDERSNO VAULT ACCESS
04 / RECONSTRUCTED
Waiting for authorized reconstruction…
LOCAL STITCH0 MAPPINGS USED
04

THE PRIVACY COMPILER

We don't simply hide names.
We determine what the task needs.

Detection is a component. The product is the decision over relevance, precision, linkability and the minimum useful representation.

ORIGINAL FACTCLASSCOMPILER ACTIONEXTERNAL REPRESENTATION
John SmithPERSONPSEUDONYMIZE<PERSON_01>
Rijksmuseum TwentheORGANISATIONCONCEPTUALIZEregional cultural museum
€82,471FINANCEGENERALIZE€80K–€90K
mira@company.exampleEMAILREMOVE—
API credentialSECRETBLOCK<API_SECRET_PRESENT>
INPUTDOCUMENT + TASK + POLICY→OUTPUTMINIMUM VIABLE CONTEXT
05

SEMANTIC COMPARTMENTS

Split by meaning.
Route by risk.

SHARD separates a task into coherent compartments. Each route receives only its required facts; synthesis happens locally.

PRECISE CLAIM Exposure reduction, not cryptographic secrecy. Correlation remains possible when providers or fragments overlap.

DOCUMENTACQUISITION AGREEMENT
AOBLIGATIONS→MODEL A
BFINANCIAL TERMS→MODEL B
CTERMINATION→LOCAL
LOCAL SYNTHESIS ✓
06

PRIVACY RECEIPTS

Proof of what happened.
Not a copy of the data.

Every protected operation records policy, provider, transformations, exposure reduction and a verifiable hash—without recording the prompt or mapping.

  • Operational observability
  • Procurement evidence
  • Debugging without raw content
  • Optional onchain hash payload
S/PRIVACY RECEIPT
VERIFIED LOCAL
REQUESTshr_8491029
ProviderConfigured upstreamPolicyEU-Enterprise-v2Original context18,421 tokensExternal context9,713 tokens
CONTEXT REDUCED47.3%
Sensitive entities42
Pseudonymized21
Generalized08
Conceptualized06
Removed07
Credentials blocked00
RISKHIGH→LOW
RECEIPT HASHsha256:8a492d10…4e7b

Receipt contains transformation metadata only. No prompt, identity, or Vault mapping.

07

OPENAI-COMPATIBLE GATEWAY

Change one endpoint.
Protect every request.

Self-host SHREDDAR inside your boundary, point an OpenAI-compatible client at it, and inspect a Privacy Receipt on every response.

READ THE ALPHA DOCS →
terminal
# run inside your trusted boundary
export OPENAI_BASE_URL=http://127.0.0.1:8787/v1

# your existing client
client.chat.completions.create(
  model="configured-upstream",
  messages=[...],
  shreddar_task="Review obligations"
)
RESPONSESHREDDAR RISK: LOW
08

AGENTS + MCP

Mediate every boundary.
Not just the first prompt.

MODELprompt / response→
SHREDDARinspect / transform→
AGENTplan / invoke→
SHREDDARauthorize / minimize→
MCP TOOLbounded context
09 / SECURITY

Fail closed.
Keep mappings separate.

No secrets in prompts. No arbitrary Vault lookup. No raw-content telemetry. Returned output is inspected before local reconstruction.

READ THE THREAT MODEL →
10 / PAY PER USE

Privacy infrastructure
agents can buy per request.

Strict PAYG only: one operation, one posted price, one completed output. The alpha includes the adapter boundary; live x402 settlement stays held until route verification.

FIAT / API + x402-READY

PRIVATE INFORMATION STAYS PRIVATE.
USEFUL MEANING CAN TRAVEL.

SHRED YOUR
FIRST REQUEST.

OPEN THE PLAYGROUND ↗VIEW THE CONSOLE →