⚡ Quickstart

CMMC gateway in 60 seconds

HoundShield is an OpenAI-compatible proxy. Every AI query from your team passes through it first. CUI, CAGE codes, contract numbers, PHI and secrets are detected and blocked before they reach any provider.

Get your keys

Point your SDK at HoundShield

Change the baseURL in any OpenAI-compatible client to the gateway. Zero behavior change for your team.

Every query is now CMMC-monitored

CUI, CAGE codes, contract numbers, PHI and secrets are flagged and blocked in <10ms — before anything leaves your perimeter.

Your gateway URL

https://proxy.houndshield.com/v1

Authentication

Two keys, two scopes

The gateway endpoint uses your normal provider key as a Bearer token — you keep using ChatGPT/Copilot/Claude exactly as before. The management endpoints (audit log, quarantine, policy) require an admin token sent as the x-admin-token header, so no one with mere network reach can release quarantined CUI or weaken your policy.

Mode B for CUI. The hosted gateway (https://proxy.houndshield.com) is for demos and non-CUI evaluation. For CUI workloads run the identical API on your own Docker host — the base URL is the only change, and prompt content never leaves your boundary.

Detection

What gets detected

CUIFOUO markings, CAGE codes, DoD contract numbers, clearance/ITAR language.

PHIAll 18 HIPAA identifiers — MRN, DOB, SSN, insurance IDs, diagnoses.

PIISSN, addresses, phone, email, financial account numbers.

SecretsAPI keys, cloud credentials, tokens, private keys in prompt bodies.

Source & IPProprietary source code and internal identifiers.

Reference

API reference

POST/v1/chat/completions

Gateway intercept

The main endpoint. OpenAI-compatible chat completions that pass through HoundShield first — every message is scanned for CUI, CAGE codes, contract numbers, PHI/PII and secrets in <10ms before the request reaches the upstream provider. Point any OpenAI-compatible SDK at the gateway and your code is unchanged.

Auth · Bearer <your provider key> — the proxy validates its own license server-side.

Headers

AuthorizationstringrequiredBearer <provider API key> (e.g. your OpenAI key).

X-ProviderstringUpstream provider: openai (default), anthropic, openrouter.

X-Provider-Api-KeystringOverride the upstream key per request.

X-User-IdstringAttribute the event to a user (audit trail).

X-Session-IdstringGroup events into a session.

Request · cURL

curl https://proxy.houndshield.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-User-Id: alex@acme-defense.com" \
  -d '{
    "model": "gpt-4o",
    "messages": [
      { "role": "user", "content": "Summarize our CAGE 1ABC2 contract" }
    ]
  }'

Response

// When the prompt is CLEAN, you get the normal OpenAI response.
// When a violation is detected, the request is blocked locally:
{
  "id": "chatcmpl-hs-9f3c1a",
  "object": "chat.completion",
  "houndshield": {
    "action": "BLOCKED",
    "risk_level": "CRITICAL",
    "detections": [
      { "engine": "CUI", "pattern": "CAGE code", "nist_control": "SC.L2-3.13.1" }
    ],
    "message": "Blocked before leaving your perimeter. Logged to the audit chain."
  }
}

GET/health

Health check

Liveness probe. No auth — safe for load balancers and uptime monitors.

Auth · None.

Request · cURL

curl https://proxy.houndshield.com/health

Response

{ "status": "ok", "version": "2.0.0", "source": "houndshield-proxy", "ooda": true }

GET/v1/stats

Live stats

Aggregate counters for the gateway — total scanned, blocked, quarantined, and average detection latency.

Auth · None (aggregate counts only, no prompt content).

Request · cURL

curl https://proxy.houndshield.com/v1/stats

Response

{ "success": true, "data": { "scanned": 14388, "blocked": 512, "quarantined": 37, "avg_ms": 8 } }

GET/v1/events admin token

Audit log

The tamper-evident, SHA-256 hash-chained event log — the C3PAO evidence trail. Filter by action or time window and page through with limit/offset.

Auth · Admin token.

Parameters

limitintMax rows (default 100, max 500).

offsetintPagination offset.

actionstringFilter: BLOCKED | QUARANTINED | ALLOWED.

sinceISO 8601Only events at/after this timestamp.

Request · cURL

curl "https://proxy.houndshield.com/v1/events?action=BLOCKED&limit=50" \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN"

Response

{
  "success": true,
  "data": [
    {
      "request_id": "9f3c1a…",
      "action": "BLOCKED",
      "risk_level": "CRITICAL",
      "engine": "CUI",
      "nist_control": "SC.L2-3.13.1",
      "hash": "a3f1…",
      "prev_hash": "77c8…",
      "ts": "2026-07-18T14:22:33Z"
    }
  ]
}

GET/v1/quarantine admin token

Quarantine queue

List requests held for human review. Scoped to one org; filter by status.

Auth · Admin token.

Parameters

org_idstringrequiredOrganization to scope to.

statusstringpending (default) | released | blocked.

limitintMax rows (default 100, max 500).

Request · cURL

curl "https://proxy.houndshield.com/v1/quarantine?org_id=acme-defense&status=pending" \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN"

Response

{ "success": true, "data": [ { "request_id": "9f3c1a…", "risk_level": "HIGH", "engine": "PHI", "ts": "2026-07-18T14:20:01Z" } ] }

PUT/v1/quarantine/:requestId admin token

Review a quarantined request

Release (allow through) or block a held request. The decision and reviewer are written to the audit chain.

Auth · Admin token.

Parameters

statusstringrequiredreleased | blocked (body).

reviewed_bystringWho made the call (body).

Request · cURL

curl -X PUT https://proxy.houndshield.com/v1/quarantine/9f3c1a \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "status": "released", "reviewed_by": "alex@acme-defense.com" }'

Response

{ "success": true }

PUT/v1/policy/:orgId admin token

Update org policy

Read (GET) or update (PUT) an organization's detection policy — which engines are active and how each risk level is handled (block vs quarantine).

Auth · Admin token.

Request · cURL

curl -X PUT https://proxy.houndshield.com/v1/policy/acme-defense \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "engines": { "CUI": "block", "PHI": "block", "PII": "quarantine" } }'

Response

{ "success": true, "org_id": "acme-defense" }

No infrastructure change required. Your employees keep using ChatGPT, Copilot or Claude — HoundShield sits in the middle, transparently, on your hardware.

⚡ Quickstart

CMMC gateway in 60 seconds

Get your keys

Point your SDK at HoundShield

Change the baseURL in any OpenAI-compatible client to the gateway. Zero behavior change for your team.

Every query is now CMMC-monitored

CUI, CAGE codes, contract numbers, PHI and secrets are flagged and blocked in <10ms — before anything leaves your perimeter.

Your gateway URL

https://proxy.houndshield.com/v1

Authentication

Two keys, two scopes

Detection

What gets detected

CUIFOUO markings, CAGE codes, DoD contract numbers, clearance/ITAR language.

PHIAll 18 HIPAA identifiers — MRN, DOB, SSN, insurance IDs, diagnoses.

PIISSN, addresses, phone, email, financial account numbers.

SecretsAPI keys, cloud credentials, tokens, private keys in prompt bodies.

Source & IPProprietary source code and internal identifiers.

Reference

API reference

POST/v1/chat/completions

Gateway intercept

Auth · Bearer <your provider key> — the proxy validates its own license server-side.

Headers

AuthorizationstringrequiredBearer <provider API key> (e.g. your OpenAI key).

X-ProviderstringUpstream provider: openai (default), anthropic, openrouter.

X-Provider-Api-KeystringOverride the upstream key per request.

X-User-IdstringAttribute the event to a user (audit trail).

X-Session-IdstringGroup events into a session.

Request · cURL

curl https://proxy.houndshield.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-User-Id: alex@acme-defense.com" \
  -d '{
    "model": "gpt-4o",
    "messages": [
      { "role": "user", "content": "Summarize our CAGE 1ABC2 contract" }
    ]
  }'

Response

// When the prompt is CLEAN, you get the normal OpenAI response.
// When a violation is detected, the request is blocked locally:
{
  "id": "chatcmpl-hs-9f3c1a",
  "object": "chat.completion",
  "houndshield": {
    "action": "BLOCKED",
    "risk_level": "CRITICAL",
    "detections": [
      { "engine": "CUI", "pattern": "CAGE code", "nist_control": "SC.L2-3.13.1" }
    ],
    "message": "Blocked before leaving your perimeter. Logged to the audit chain."
  }
}

GET/health

Health check

Liveness probe. No auth — safe for load balancers and uptime monitors.

Auth · None.

Request · cURL

curl https://proxy.houndshield.com/health

Response

{ "status": "ok", "version": "2.0.0", "source": "houndshield-proxy", "ooda": true }

GET/v1/stats

Live stats

Aggregate counters for the gateway — total scanned, blocked, quarantined, and average detection latency.

Auth · None (aggregate counts only, no prompt content).

Request · cURL

curl https://proxy.houndshield.com/v1/stats

Response

{ "success": true, "data": { "scanned": 14388, "blocked": 512, "quarantined": 37, "avg_ms": 8 } }

GET/v1/events admin token

Audit log

The tamper-evident, SHA-256 hash-chained event log — the C3PAO evidence trail. Filter by action or time window and page through with limit/offset.

Auth · Admin token.

Parameters

limitintMax rows (default 100, max 500).

offsetintPagination offset.

actionstringFilter: BLOCKED | QUARANTINED | ALLOWED.

sinceISO 8601Only events at/after this timestamp.

Request · cURL

curl "https://proxy.houndshield.com/v1/events?action=BLOCKED&limit=50" \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN"

Response

{
  "success": true,
  "data": [
    {
      "request_id": "9f3c1a…",
      "action": "BLOCKED",
      "risk_level": "CRITICAL",
      "engine": "CUI",
      "nist_control": "SC.L2-3.13.1",
      "hash": "a3f1…",
      "prev_hash": "77c8…",
      "ts": "2026-07-18T14:22:33Z"
    }
  ]
}

GET/v1/quarantine admin token

Quarantine queue

List requests held for human review. Scoped to one org; filter by status.

Auth · Admin token.

Parameters

org_idstringrequiredOrganization to scope to.

statusstringpending (default) | released | blocked.

limitintMax rows (default 100, max 500).

Request · cURL

curl "https://proxy.houndshield.com/v1/quarantine?org_id=acme-defense&status=pending" \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN"

Response

{ "success": true, "data": [ { "request_id": "9f3c1a…", "risk_level": "HIGH", "engine": "PHI", "ts": "2026-07-18T14:20:01Z" } ] }

PUT/v1/quarantine/:requestId admin token

Review a quarantined request

Release (allow through) or block a held request. The decision and reviewer are written to the audit chain.

Auth · Admin token.

Parameters

statusstringrequiredreleased | blocked (body).

reviewed_bystringWho made the call (body).

Request · cURL

curl -X PUT https://proxy.houndshield.com/v1/quarantine/9f3c1a \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "status": "released", "reviewed_by": "alex@acme-defense.com" }'

Response

{ "success": true }

PUT/v1/policy/:orgId admin token

Update org policy

Read (GET) or update (PUT) an organization's detection policy — which engines are active and how each risk level is handled (block vs quarantine).

Auth · Admin token.

Request · cURL

curl -X PUT https://proxy.houndshield.com/v1/policy/acme-defense \
  -H "x-admin-token: $HOUNDSHIELD_ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "engines": { "CUI": "block", "PHI": "block", "PII": "quarantine" } }'

Response

{ "success": true, "org_id": "acme-defense" }

No infrastructure change required. Your employees keep using ChatGPT, Copilot or Claude — HoundShield sits in the middle, transparently, on your hardware.