auditlog-ai 0.1.0

Automatic compliance logging for AI agents. One line install, vendor-agnostic, audit-ready.

AgentLog

Audit trail for AI agents. One decorator, zero dependencies.

pip install auditlog-ai

The problem

Your agents call models, execute tools, make decisions. If something goes wrong — or an auditor asks what happened — you have nothing. No logs, no trail, no proof.

AgentLog fixes that. Every agent action gets a structured event: who did what, when, which model, who authorized it. Append-only JSONL, SHA-256 hashed I/O, ready to ship to any backend.

30 seconds to production

Decorator — wrap any function, done:

from agentlog import log_agent

@log_agent(agent_id="support-bot", model_id="gpt-4")
def answer(prompt: str) -> str:
    return call_my_model(prompt)

answer("How do I reset my password?")
# → agentlog.jsonl gets a new line

Context manager — when you need timing:

from agentlog import AgentLogger

with AgentLogger(agent_id="support-bot", model_id="gpt-4") as logger:
    result = call_my_model("How do I reset my password?")
    logger.set_output(result)
# → logged with latency_ms

Direct call — full control:

from agentlog import AgentLogger

logger = AgentLogger(agent_id="support-bot", model_id="gpt-4")
logger.log(input_text="Hello", output_text="Hi there")

What gets logged

{
  "event_id": "a1b2c3d4-...",
  "agent_id": "support-bot",
  "session_id": "e5f6g7h8-...",
  "timestamp": "2026-03-30T12:00:00+00:00",
  "input_hash": "2cf24b72...",
  "output_hash": "486ea46...",
  "model_id": "gpt-4",
  "authorized_by": "system",
  "latency_ms": 423
}

Inputs and outputs are SHA-256 hashed. The actual content never touches the log unless you put it in metadata.

Bring your own backend

JSONL by default. Supabase if you want cloud:

from agentlog import AgentLogger, SupabaseBackend

backend = SupabaseBackend(url="https://xxx.supabase.co", key="your-key")
logger = AgentLogger(agent_id="my-agent", model_id="gpt-4", backend=backend)

pip install auditlog-ai[supabase]

Writing your own backend is one method: save(event: AgentEvent).

Design decisions

• Zero dependencies. Core uses only the standard library. No vendor lock-in.
• Hashed by default. Input/output never stored in plaintext unless you opt in.
• Append-only JSONL. Grep it, stream it, pipe it. No database required.
• Vendor-agnostic. Works with OpenAI, Anthropic, LangChain, your custom stack.

License

MIT