Auditant records what an agent attempted, which rule decided, who approved it and what happened next — then seals the record so an auditor can check it without trusting you, or us. For the person who signs; connected by the person who ships, in two lines.
● Our own record: 915 actions, sealed through #914, 71 of 73 seals countersigned by an independent clock. See it
Priced per agent, never per event. Free for one agent. Only fingerprints leave your environment — never the contents.
A real receipt, issued by Auditant on a sample record. Anyone can verify it offline with the published key.
LiteLLMVapiOpenTelemetryAmazon S3A staged rollout that earns the right to enforce before it interrupts production.
Use the SDK, your gateway, or the OpenTelemetry pipeline you already run. The record starts; nothing is blocked.
Model calls, tool calls, cost, people and resulting effects land in one ordered record, with coverage gaps stated separately.
Write limits in plain English. Every rule starts by observing, then holds or refuses only after you arm it.
Export one evidence bundle or examiner packet. The reviewer verifies it offline without an Auditant account.
Native instrumentors where they exist; open protocols and explicit event APIs everywhere else. The connection path is stated on every group.
OpenAI, Anthropic, Gemini, any tool-calling LLM
Automatic for the named SDKs; every other provider connects through OpenTelemetry, LiteLLM, or the event API.
LangChain, LangGraph, CrewAI, LlamaIndex, OpenAI Agents SDK, your own loop
Framework spans are captured automatically when installed; a custom loop uses record() and decide().
Retell, Vapi, any audio loop
Vapi connects directly. Retell and other call stacks use a verified webhook relay or the canonical voice ingest API.
MCP servers, REST, Python & TypeScript SDKs, internal services
Record the attempted call at the agent, then join it to the effect your service or gateway reports.
LiteLLM, OpenTelemetry, Vercel AI SDK, any OTLP collector
Keep the pipeline you already run; Auditant maps its spans into the same evidence format.
Amazon S3 Object Lock, JSON bundles, examiner packets, SIEM webhooks
The record leaves as portable evidence, not a screenshot that only works inside our account.
A $50,000 payout: recorded, held, approved by name, executed, sealed. Every block carries the fingerprint of the one before it, so nothing can be quietly changed later.
Which agent acted, in which session, on whose behalf. Without that last part there is no per-customer answer, complaint or erasure later.
A fingerprint of what went in and what came out, the model, and the price. The prompt itself never leaves your environment.
“Moving more than the threshold requires a named human to approve first.” Synchronous, on the request path — the side effect has not happened. A signed receipt is issued for the decision itself.
Taken from her authenticated server session, not a service account — the field every investigation ends on.
The effect plane confirms it landed — the ledger entry, not the agent's claim that it made one.
A checkpoint signs the chain head with our key, an RFC 3161 timestamp authority countersigns it, and a copy is held in write-once storage. From here, changing any earlier row is a computable fact.
Start with what needs attention. Open the action to see the rule, the evidence and the accountable human response, then follow the same record through activity, coverage and export.
Two people come to this page — the one who signs and the one who ships. The same answers, side by side.
Every action your agents take — what they read, what they decided, who approved, what actually happened — written as it happens, in an order nobody can quietly rearrange.
Model calls, tool calls, rule decisions, human approvals and the effects that landed, in one format. Each event carries a fingerprint of the previous one, on a store that only appends.
Anything you name in plain English. Above a threshold, a person you name has to say yes first, and their yes is on the record under their name.
One synchronous check before the action runs: allow, refuse, change, hold or defer. Every rule starts in watch-only mode. A halt switch beats every rule and is itself recorded.
Fingerprints, not contents. We can prove nothing was altered without ever seeing a prompt, a document or a customer's data.
By default: who, what, when, the outcome, and a fingerprint of every input and output. Optionally the contents go to your own storage bucket. Never health or card data.
Sealed by us, countersigned by an independent clock, and checkable by the auditor on their own laptop — without asking us anything.
Signed seals, an independent timestamp on each, and an optional write-once copy. One command — auditant verify — recomputes every fingerprint and names the first row that no longer matches.
Four rows of a real sample record. Change one; the verifier tells you exactly where the record stopped being true. The real verifier runs on your own machine — one command.
Untouched, the record verifies. Change what row 006 says happened and that row — and every row after it — stops matching the seal.
Open only what you need. The product details stay available to readers and search engines without turning the page into a wall of copy.
It is a chronological record of an agent’s model calls, tool calls, policy decisions, human approvals and resulting actions. Auditant links those events into one tamper-evident chain so an examiner can follow a decision from intent to outcome.
Both. Auditant records every observed action and can evaluate policy before a tool runs. A rule can allow the action, refuse it, or hold it until an authenticated human approves or rejects it.
Verification proves that the records Auditant received have not been altered after they were sealed. It does not prove that every action reached the record; deployment coverage is measured and reported separately.
No. The default record stores metadata and cryptographic hashes rather than prompt or response payloads. You can prove integrity while keeping sensitive content in storage you control.
OpenAI, Anthropic and Gemini model calls can be auto-instrumented. LangChain, LangGraph, CrewAI, LlamaIndex and the OpenAI Agents SDK connect through their framework instrumentation; any other tool-calling model or custom loop can emit the same evidence through OpenTelemetry, LiteLLM, the Python or TypeScript SDK, or the event API.
Pricing is per agent, never per recorded event. One agent is free to evaluate, and paid bands include unlimited recorded actions so cost never rewards incomplete coverage.
