InficyFrequently asked questions
Using Inficy
What does Inficy show me?
What your AI agents are actually doing. Each execution becomes a readable activity record: where the agent went, which tools, models and services it used, how long it took, what happened when it failed or recovered, and who intervened.
What agent activity can it capture?
Ordered actions with timestamps: tool calls, model calls, web requests, file writes, delegations to other services, human approvals, errors, retries and final outcomes.
Can I see tools, websites, time, spend, failures and handoffs?
Yes, where the agent reports them. Time is shown as wall-clock duration and, separately, as active execution time. Financial activity is shown in the state it was observed in: requested, attempted, approved, succeeded, failed or refunded.
Does Inficy capture everything?
No. Inficy shows what the agent reported through the capture integration. Actions taken outside the instrumented execution loop, or details a model or provider does not expose, are not visible.
What happens when activity is not observable?
It is left out rather than guessed. Inficy does not infer a successful outcome from a request, and does not present unreported steps as if they happened.
Connecting agents
How do I connect an agent?
Create an agent connection in your workspace, copy the one-time setup code, then run the installer or connector for your runtime. It exchanges the code for a workspace-scoped reporting token, configures capture and runs a self-check. The agent reaches Ready only after a real test execution has been received and verified. The one-command installer currently covers Hermes; the OpenAI Agents SDK connector is installed as a package.
Which runtimes and frameworks are supported?
Two runtimes produce certifiable, runtime-observed evidence today: Hermes Agent (available, one-command installer) and the OpenAI Agents SDK for TypeScript (beta). ChatGPT connects over MCP and produces agent-reported session records, which are never certifiable. No other runtime is supported, and Inficy does not claim to instrument runtimes it has not been tested against.
Does Inficy need my model-provider API keys?
No. Inficy never asks for or stores your model-provider credentials. The agent reports its own execution activity using a workspace-scoped reporting token.
Execution records
What is an execution record?
One end-to-end thing an agent did, stored as an ordered sequence of actions plus its outcome. It is the receipt for that execution: readable by a person and structured enough to export.
What makes a record verifiable?
Executions are recorded as canonical NexArt artifacts and verified offline with the NexArt SDK before they are trusted in the timeline. A record is therefore tamper-evident: an altered artifact fails verification instead of being silently accepted.
Can I export an execution?
Yes. Executions and their evidence can be exported, and anyone holding the export can verify it offline without access to your workspace.
What is a confidential execution?
An execution whose sensitive payloads are sealed: the artifact commits to the content without disclosing it, and the opening material remains with you as the data owner. Verification still works while the payload stays private.
Certification
Do I have to certify executions?
No. Recording, inspection and export are included with every workspace. Certification is optional, deliberate and paid per execution using certification credits.
What does certification add?
An independent attestation from the NexArt Node that a specific recorded execution existed in exactly that form at that time and has not been altered since. It does not prove the agent's decision was correct, safe or compliant, and it says nothing about actions that were never captured.
Is recorded time different from certification time?
Yes. The recorded time is when the execution happened; the certification time is when the artifact was attested. The two are shown separately and should never be read as the same event.
Security and privacy
How are agent credentials protected?
Reporting tokens and machine credentials are stored only as keyed digests, are bound to one workspace and an explicit scope set, and can be revoked at any time.
How is my data isolated from other customers?
Every tenant-owned table is protected by Postgres row-level security scoped to workspace membership, so a query that omits its workspace scope returns nothing rather than another tenant's data.
How long is data retained?
Retention depends on your plan and is enforced in the database. Pilot workspaces can request export or deletion of their workspace data at any time by contacting support.
Who else processes my data?
The subprocessors listed on the subprocessors page. Nothing is submitted to the NexArt Node unless you explicitly request certification of a specific execution.