Skip to main content

Changelog

All notable changes to the Auditail SDKs (auditail-sdk on npm, auditail on PyPI) and the Auditail server. Dates are release dates. The server is deployed continuously; server-only changes are listed under the release they shipped with.

0.9.0 (2026-11-16)​

  • Plan before acting: simulate() (a dry run that records and reserves nothing) and getCapabilities() / get_capabilities() (allowed actions, approval conditions, remaining budget). A policy can turn capabilities off with "introspection": false.
  • Pause and resume for workflow engines: correlationId and a signed per-request callbackUrl, retried on failure, plus resume(). Examples for Temporal, Inngest and LangGraph.
  • Idempotent decisions: every guard() call with serverPolicy sends an idempotency key, so retries can't open a second approval or count usage twice. Pass your own idempotencyKey for retries across restarts.
  • Outcomes: reportOutcome() / report_outcome() records settled, failed, reversed, refunded or voided, with links to compensating actions. Trusted failed and voided reports (approver token, or Airwallex and Modern Treasury events) give limit usage back; agent reports are recorded only.
  • Session taint: markUntrusted() / mark_untrusted() and sessionId. After a session reads untrusted content, money-moving actions in it need a human (policy taintedSession). Only approvers can clear it.
  • MCP proxy: auditail-mcp-proxy decides every MCP tools/call before it reaches the server, with per-tool action mapping and untrusted-content tools. McpGuard for your own MCP server.
  • Act-only agent keys. An agent key can be restricted so it can act (decisions, approvals, kill switch, events) but can't read the audit log, exports, reports, policy or connections.
  • Measured per-account throughput and contention behavior, documented on the security page.
  • Unknown pages on auditail.io return a proper 404.

0.8.1 (2026-11-15)​

  • Package metadata: the publisher is Auditail ([email protected]). Examples use auditail consistently. No behavior changes.

0.8.0 (2026-11-15)​

  • Payment provider integrations. Stripe Issuing real-time card authorizations, enforced by your policy. Airwallex and Modern Treasury approval webhooks, with Auditail's verdict for reviewers and flagged overrides. No payment credentials are shared.
  • checkReceipt() / check_receipt(): verify and redeem a receipt in one call.
  • firstTime rules accept a known list of payees set up outside the agent.

0.7.0 (2026-11-15)​

  • Two-person approval (approvers: 2, or requiredApprovals per request).
  • First-time rules for new beneficiaries or accounts.
  • Conditional limits (when), e.g. separate caps per currency.
  • Single-use receipts: redeemReceipt() / redeem_receipt(), with safe retries.
  • Checkpoint anchoring: signed checkpoints pushed to storage you control.
  • The verifier now reports a log cut short against a saved checkpoint.
  • A self-hosting guide, tested from a clean copy of the repository.

0.6.1 (2026-11-14)​

  • Package metadata and README updates. No code changes.

0.6.0 (2026-11-14)​

  • Kill switch: freeze all agents, one agent or one action. Only approvers can lift a freeze.
  • Shadow mode and policy backtests against recorded traffic.
  • Signed examiner report answering the four questions supervisors ask about AI agents.

0.5.0 (2026-11-14)​

  • Server-side policies: default-deny allowlists, atomic spending and rate limits, untrusted-source rules, and reason codes with hints the agent can act on.

0.4.0 (2026-11-14)​

  • Signed receipts, signed checkpoints and offline-verifiable evidence exports (auditail-verify).

0.3.1 (2026-11-14)​

  • Fix: a rule that throws an error now blocks immediately (fails closed) and never asks a human.

0.3.0 (2026-11-14)​

  • Human approval: approver tokens, dashboard decisions and signed webhooks.

0.2.0 (2026-11-13)​

  • First public beta: SDKs for TypeScript and Python, hosted hash-chained audit log and dashboard.