yapguard

Privacy-first, safety evaluation API for LLM agents.

A lightweight guard rail you drop in front of any LLM pipeline. Evaluates messages before they reach the model and before output reaches users, blocking prompt injection, PII leakage, credential exposure, dangerous tool calls, and more.

Start in 3 steps

• Register once via proof-of-work and store your API key
• Call POST /v1/evaluate before executing tool calls
• If safe=false, halt that step and surface reasons

Then expand to full loop coverage

• message_type=user before model input
• message_type=tool_call before tool execution
• message_type=tool_result before returning tool output to the model
• message_type=assistant before returning final output to the caller

What it checks

• Prompt injection & malicious intent (ML classifier)
• PII in user input — email, SSN, credit card, phone
• Dangerous tool calls — shell commands, SQL injection, blacklisted domains, internal network access
• Redirect chains that resolve to internal destinations
• System prompt leakage in assistant output
• Secrets and credentials in tool results and LLM output

API

Full OpenAPI spec →

Concrete example (/v1/evaluate)

Tool-call-only evaluation with one allowed request and one blocked request:

# allowed: external URL, no dangerous patterns
curl -s -X POST https://yapguard.com/v1/evaluate \
  -H "Authorization: Bearer $YAPGUARD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message_type": "tool_call",
    "message": "{\"tool\":\"webfetch\",\"args\":{\"url\":\"https://example.com/docs\"}}"
  }' | jq .

# allowed response
{
  "safe": true,
  "risk_score": 0.0,
  "reasons": []
}

# blocked: localhost destination (internal network)
curl -s -X POST https://yapguard.com/v1/evaluate \
  -H "Authorization: Bearer $YAPGUARD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message_type": "tool_call",
    "message": "{\"tool\":\"webfetch\",\"args\":{\"url\":\"http://127.0.0.1:8080/admin\"}}"
  }' | jq .

# blocked response
{
  "safe": false,
  "risk_score": 1.0,
  "reasons": [
    {
      "rule_id": "tool_call.internal_network_access",
      "severity": "high",
      "detail": "tool call targets internal/local destination(s): 127.0.0.1"
    }
  ]
}

If blocked, do not execute the tool call. Return reasons to the caller or policy handler.

Integrate using Claude Code

If you're a developer integrating via an AI coding assistant, point it at the spec and it will wire up the guard automatically. Example prompt:

# paste into Claude Code

Fetch https://yapguard.com/openapi.json and integrate yapguard into
this codebase as a safety guard rail. Register for an API key using
the proof-of-work flow, store it in .env as YAPGUARD_API_KEY, then
wrap every LLM call so that:
  1. user input is evaluated before being sent to the model
  2. tool calls are evaluated before execution
  3. tool results and assistant output are evaluated before returning
     to the caller
Halt the pipeline and surface the reasons array if safe=false.

Self-registration

No account, no email. Agents register autonomously via proof-of-work: a short compute challenge that takes a few seconds to solve and serves as a lightweight spam deterrent. The API key is returned exactly once — store it.

# 1. get a challenge
curl -s -X POST https://yapguard.com/v1/register/challenge \
  -H "Content-Type: application/json" | jq .

# 2. solve the proof-of-work challenge, then submit
curl -s -X POST https://yapguard.com/v1/register/solve \
  -H "Content-Type: application/json" \
  -d '{"challenge_id":"<id>","nonce":"<nonce>"}' | jq .

level=info method=POST path=/v1/evaluate status=200 duration_ms=23 remote_addr=203.0.113.42 country=US message_type=tool_call tool_name="bash" input_bytes=1024 safe=true risk_score=0.0000 reason_ids=none
level=info method=POST path=/v1/evaluate status=400 duration_ms=18 remote_addr=203.0.113.42 country=US message_type=user tool_name="na" input_bytes=312 safe=false risk_score=0.9400 reason_ids=prompt_injection,credential_exposure

openapi.json  ·  llms.txt