Skip to main content

auditail (Python SDK)

Controls and evidence for AI agents that move money. Limits agents can't change, human approval, a kill switch, and a signed, tamper-evident record of every decision.

📖 Full Documentation · 🚀 Join the Private Beta


Install

pip install auditail
pip install "auditail[verify]"   # adds verify_receipt() for the service that executes actions

Quickstart: limits and approval for a payout tool

An approver saves a policy once (dashboard or PUT /v1/policy); the agent can't change it:

{
  "defaultAction": "deny",
  "actions": {
    "create_payout": {
      "requireApproval": [{ "field": "amount", "op": ">", "value": 10000 }],
      "limits": [{ "name": "daily-total", "window": "24h", "sum": "amount", "max": 25000, "onExceed": "deny" }]
    }
  }
}

The agent checks every payout before it runs:

import os
from auditail import Auditail

auditail = Auditail(api_key=os.environ["AUDITAIL_API_KEY"], server_policy=True)

payout = {"amount": 9999, "beneficiary": {"id": "ben_7"}}
result = auditail.guard(action="create_payout", payload=payout, agent_id="payout-agent")

if result.approved:
    payments.create_payout(payout, auditail_receipt=result.receipt)  # your payment client
else:
    print(result.code, result.reason)
    # Third $9,999 payout today:
    # limit_exceeded Denied: limit "daily-total" would be exceeded (29,997 of 25,000 in 24h).

Payouts over $10,000 wait for a human to approve in the dashboard or via your webhook. Any other action is refused. Every decision goes into a tamper-evident audit log.

Agent frameworks

Call guard() at the top of the tool and return the reason to the model when it says no:

Private beta

The SDK, API and dashboard are live. Guardrail rules run locally in your process; every decision is delivered in the background to a hash-chained audit log you can browse and verify at app.auditail.io.

  • Batched, retried, idempotent delivery of events. With local rules only, guard() doesn't wait on the network and a delivery failure never reaches your agent. With server_policy, guard() asks the server and fails closed if it can't be reached.
  • Secrets are redacted before an event leaves your process (password, token, apiKey, ... plus your own redact_keys).
  • Call auditail.flush() before a serverless function returns or the process exits.

Request an API key: keys are issued to design partners personally.

Human approval is live: a require_approval rule pauses the agent until someone approves or denies it in the dashboard or through your webhook (Slack, Teams, Discord, any HTTPS endpoint). No answer means denied. See the approval guide.

Server-side policies: allowed actions with everything else denied, spending and rate limits counted atomically on the server, rules for data from untrusted sources, and reason codes with hints the agent can act on. Approvers write them; agents can't change them. See policies.

Signed receipts: every human approval returns a receipt your payment or email service can verify before acting (verify_receipt), so an agent cannot skip the check. Logs export to a file anyone can verify offline: npx -p auditail-sdk auditail-verify. See receipts and evidence.

Kill switch, shadow mode and examiner reports: freeze every agent, one agent or one action; try a policy in shadow mode or backtest it on recorded traffic; download a signed report for your examiners. See kill switch and reports.

Roadmap (not available yet): native Slack and PagerDuty apps, private-cloud deployment.

License

MIT

Metadata

Release files for auditail 0.8.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for auditail 0.8.0
File Size Uploaded Provenance Actions
auditail-0.8.0.tar.gz 22.3 kB Without provenance Details for auditail-0.8.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for auditail 0.8.0
File Interpreter ABI Platform Actions
auditail-0.8.0-py3-none-any.whl Python 3 none any Details for auditail-0.8.0-py3-none-any.whl

Total release size: 41.6 kB

Release files / auditail-0.8.0.tar.gz

Download URL auditail-0.8.0.tar.gz
Size 22.3 kB
Tags Source
SHA-256 checksum
How to use checksums
a1c6dcf7079c76d3414bf4c5b9627202fd94d40ee30dd0f9116b879695b8cd68
BLAKE2b-256 checksum
How to use checksums
06033cc508578559a8e97159dbd8afee99c6b34dec8d851dce41ee35869db21a
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Release files / auditail-0.8.0-py3-none-any.whl

Download URL auditail-0.8.0-py3-none-any.whl
Size 19.4 kB
Tags Python 3
SHA-256 checksum
How to use checksums
a59d43cdd9acd332229420afe19b449590d0fb1ab62b0f853e3ec14ec482921d
BLAKE2b-256 checksum
How to use checksums
4e48723ec3f72ef76b8c423227285695bcffee48e4fcffe6b5dd8f8f91b363c9
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Security for auditail 0.8.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page