Promptrak Python SDK

Secure and govern your AI agents — prompt security, PII vault, and tool authorization in three lines of code.

pip install promptrak

Quickstart

from promptrak import Promptrak

client = Promptrak(api_key="tp_...", tenant_id="acme")

Tier 1 — Prompt Security (10 minutes)

Protects against prompt injection, jailbreak attempts, credential extraction, and policy bypass. Secures what goes into the model.

decision = client.evaluate(prompt, actor_id="agent-1")

if decision.action == "allow":
    response = my_llm.complete(prompt)
elif decision.action == "minimize":
    # Use the sanitized version — unsafe content has been removed
    response = my_llm.complete(decision.safe_prompt)
else:
    # "block" — do not proceed
    return "Request blocked by security policy"

decision.policy_trace contains the human-readable reason for any non-allow decision. Use it for audit logs and user-facing explanations.

Tier 2 — PII Vault (1 hour)

Protects against PII leakage through the LLM. Real values are stored in the vault — the model only sees placeholders. Output is scrubbed before it reaches the user.

safe = client.sanitize(prompt, actor_id="agent-1")
response = my_llm.complete(safe.safe_prompt)   # model sees [NAME], [EMAIL] etc.
clean = client.scrub_output(response, session_id=safe.session_id)
return clean   # safe to deliver to the user

The session links the sanitize and scrub calls. You never need to track the vault mapping — the SDK handles it.

Tier 3 — Tool Authorization (half day)

Protects against unauthorized tool execution, high-blast-radius actions, and malicious tool parameters. Every tool call requires explicit authorization before execution.

auth = client.authorize_tool("query_database", params, actor_id="agent-1")

if auth.action == "allow":
    result = db.execute(params)
elif auth.action == "pending_approval":
    # Queue for human review
    queue_for_review(auth.approval_id, params)
    return "Tool call queued for human approval"
else:
    # "deny"
    return f"Tool access denied: {auth.reason}"

Resolving a pending approval

# Called by your review system or human approver
client.submit_approval(
    auth.approval_id,
    "approve",           # or "deny"
    approver_id="reviewer-jane",
)

Choose Your Tier

Return Types

@dataclass
class Decision:
    action: str        # "allow" | "minimize" | "local-only" | "block"
    safe_prompt: str   # use instead of original when action == "minimize"
    risk_score: float  # 0.0 → 1.0
    policy_trace: list # who, what, why, which rule
    request_id: str    # for audit correlation

@dataclass
class SanitizedPrompt:
    safe_prompt: str   # prompt with PII replaced by vault placeholders
    session_id: str    # pass to scrub_output()
    action: str
    risk_score: float
    policy_trace: list

@dataclass
class ToolDecision:
    action: str        # "allow" | "deny" | "pending_approval"
    reason: str
    approval_id: str | None  # present when action == "pending_approval"

Error Handling

All SDK methods raise PromptrakError on failure.

from promptrak import PromptrakError

try:
    decision = client.evaluate(prompt, actor_id="agent-1")
except PromptrakError as e:
    print(e.status_code)   # HTTP status, or None for connection errors
    print(e.request_id)    # correlate with backend audit logs
    raise

Token management and retry on transient failures are handled internally. You never call /auth/token directly.

Self-hosted deployments

client = Promptrak(
    api_key="tp_...",
    tenant_id="acme",
    base_url="https://your-deployment.internal",
)

Development

pip install -e ".[dev]"
pytest