Wytness Python SDK

⚠️ Deprecated — use wytness-agt instead

The Wytness v2 wrapper SDK ships as wytness-agt, built on top of Microsoft's Agent Governance Toolkit (AGT) (MIT, released April 2026). AGT does identity Ed25519, IATP delegation, Merkle hash chain, and policy enforcement — primitives this SDK rolled itself. wytness-agt is a thin wrapper that adds the Wytness layers AGT deliberately leaves out: customer-held PII tokenisation, per-event Ed25519 envelope (non-repudiation), and bring-your-own-storage routing.

Migrate when convenient. This package (wytness-sdk) remains installable for now but is feature-frozen — no new features land here past 2026-12-31, and security-only patches are committed through 2027-06-30. After that the package goes read-only on PyPI.

Install the v2 wrapper:

pip install wytness-agt

Then call wytness_agt.init(...) with your API key + signing key + PII keypair. The new install bundle is one copy-paste block from the operator's Keys page.

See docs/agt-wrapper-design.md for the architecture, and the wytness-agt README for the new usage.

Audit logging for AI agents. Every action your AI agent takes is cryptographically signed, hash-chained, and streamed to the Wytness platform for compliance and observability.

Installation

pip install wytness-sdk

One-time setup — generate and configure your signing key

The SDK refuses to construct an AuditClient until a signing keypair is configured — it raises WytnessSetupError otherwise. This is by design: pre-configuring ensures the matching public PEM is already registered with the platform, so /ingest cannot 412 on signature verification.

1. Generate the keypair at https://app.wytness.ai/keys. The private key is created in your browser and offered for download as the literal string wyt_pk_sign_<base64> — Wytness never sees it. The matching public PEM is registered with the platform automatically.

2. Configure it for the SDK. Pick one — env var is recommended:

   Env var (recommended for every environment): set the downloaded value as the WYTNESS_SIGNING_KEY env var in your secrets manager / .env / deployment platform's env config. Since 0.12.3 the SDK reads it automatically — no boilerplate in the constructor:

   client = AuditClient(
       agent_id="my-agent",
       http_api_key=os.environ["WYTNESS_API_KEY"],
   )
   

   To pass the key explicitly instead, signing_key=os.environ["WYTNESS_SIGNING_KEY"] still works and wins precedence over the env-var auto-read.

   File on disk (local prototyping only): save the downloaded file to ./keys/signing.key. The SDK reads this path by default. ⚠️ Add keys/ to your .gitignore first — the private signing key is the root of trust for the entire audit guarantee; if it leaks into git an attacker can fabricate events that look like they came from your agent. Treat it like a database password.

3. Verify the local key matches what the portal has registered:

   wytness-show-public
   # or:
   python -m wytness.show_public
   

Quick Start

from wytness import AuditClient, audit_tool

# Initialize the client
client = AuditClient(
    agent_id="my-agent",
    human_operator_id="user-123",
    http_api_key="wyt_api_live_...",   # from app.wytness.ai settings
)

# Decorate any tool your agent calls
@audit_tool(client=client)
def send_email(to: str, subject: str, body: str):
    # Your tool implementation
    pass

# Every call is automatically logged, signed, and chained
send_email(to="team@company.com", subject="Report", body="Weekly summary")

Features

• Response capture — AI agent responses are automatically captured, truncated to 5K chars, and PII-redacted
• PII redaction — emails, SSN, TFN, credit cards, and phone numbers are automatically redacted from responses and parameter values
• Cryptographic signing — every event is signed with Ed25519
• Hash chaining — tamper-evident chain of events per agent session
• Automatic secret redaction — secrets in parameter names are automatically redacted
• HTTP transport — stream events to api.wytness.ai over HTTPS
• File fallback — events are saved locally if the API is unreachable

HTTP Emitter (Recommended)

Send events directly to the Wytness API using your API key:

client = AuditClient(
    agent_id="my-agent",
    human_operator_id="user-123",
    http_endpoint="https://api.wytness.ai",
    http_api_key="wyt_api_live_...",
)

API

AuditClient(agent_id, **kwargs)

@audit_tool(client, task_id="default", prompt=None, reasoning_summary=None, inputs_classification=None, auto_extract=True)

Decorator that wraps a function to automatically log audit events. Works with both sync and async functions. Captures:

• Function name, parameters (with secret redaction), and return value hash
• Response text (truncated to 5K chars, PII-redacted) — the function's return value as a string
• Execution duration, success/failure status, and error codes
• Cryptographic signature and hash chain link
• Per-process agent_instance_id (since 0.13.0) — UUID persisted to ./keys/.instance_id; the Wytness Agents page uses it to count distinct instances per agent_id.

Auto-captured fields by convention (since 0.13.0; disable with auto_extract=False):

• prompt — picked from any wrapped-function argument named one of user_input, prompt, question, query, input_text, message, user_message (first match, truncated to 2000 chars).
• reasoning_summary — extracted by duck-typing the return value. Anthropic-style .content text blocks and OpenAI-style .choices[0].message.content recognised, truncated to 500 chars. Plain string returns ignored (no false positives).
• inputs_classification — set to "restricted" when any kwarg name contains ssn/tfn/credit_card/card_number/passport (case-insensitive).

All three accept an explicit string OR a callable (args, kwargs) -> str that wins precedence over the convention.

client.session_id

Read-only property returning the UUID for this client instance.

client.flush(timeout=5.0)

Wait for pending HTTP requests to complete. Call before process exit in serverless/short-lived environments.

hash_value(value)

SHA-256 hash of any JSON-serialisable value. Use this when recording events manually to compute inputs_hash / outputs_hash.

from wytness import hash_value

result = my_tool(data)
outputs_hash = hash_value(result)

Logging

The SDK emits diagnostics (HTTP ingest errors, fallback replay info, key-generation warnings, record-path errors) through Python's standard logging module, under the wytness logger. Control verbosity the same way you would for any other library:

import logging

# Silence the SDK completely:
logging.getLogger("wytness").setLevel(logging.CRITICAL)

# Or just silence the info-level replay messages, keep warnings + errors:
logging.getLogger("wytness").setLevel(logging.WARNING)

# Or route everything to your own handler:
logging.getLogger("wytness").addHandler(my_handler)

If you haven't configured Python's logging at all, SDK warnings and errors appear on stderr via the default last-resort handler.

License

MIT