Frequently asked questions

Voidex is infrastructure for AI agents that take actions in business systems. It connects agents to typed tools, enforces deny-by-default policy and human approvals before anything runs, executes actions in an isolated worker, then independently checks the real system to confirm what actually happened and issues a tamper-evident receipt.

Voidex is for teams putting AI agents into real workflows: developers and automation engineers who need a controlled way for agents to call tools, operations leaders who need to see what workers did, and security or audit owners who need approvals, permissions and evidence. It is model-agnostic and not tied to one industry.

No. Voidex does not train or host a foundation model and is not a chat interface. It is the layer around agents: tool connectivity, policy, approvals, execution, verification, receipts and benchmarks. Agents from any provider can use it through the REST API, the TypeScript SDK or the MCP endpoint.

Agent frameworks help an agent decide what to do, and observability tools record what it said it did. Voidex sits on the execution path: it can block or hold an action before it happens, and after it happens it reads the target system again to confirm the effect, instead of trusting the agent’s or the API’s claim of success.

After an action executes, Voidex runs typed postcondition checks that read the target system independently, for example confirming a record exists, a field equals the requested value, line items sum to a total, or a status moved from draft to posted. Each check is passed, failed, inconclusive or not run, and only passed deterministic or human checks count as verified.

No. A model saying an action succeeded is never treated as proof. Voidex supports an optional model assessment check, but it is labelled probabilistic and can never make an action “verified” on its own; at least one deterministic or human-attested check must pass.

The action is marked failed with outcome verification_failed, even if the target API returned success. The receipt records which check failed and why, for example “Invariant broken: 799.97 ≠ 800.97”. Inconclusive checks send the action to a human review queue instead of being reported as success.

Policies can require approval for sensitive actions. An approval is bound to a hash of the exact payload, target and policy version, and it expires. If anything changes after approval, the approval no longer applies. Requesters cannot approve their own actions, and high-impact actions always need two different approvers.

If no policy rule explicitly allows an action, it is blocked. Rules match on tool, connector, risk class, environment and input values, and the most restrictive matching rule wins. Platform guardrails also apply regardless of tenant policy, such as dual approval for high-impact actions.

Only when it is safe. Voidex retries a call if the request never reached the target or the tool is declared idempotent. If a non-idempotent call’s response is lost, Voidex does not resend it; it verifies the real state and asks a human to review, so the action cannot be duplicated.

Yes. Voidex exposes a workspace’s registered tools through an MCP server endpoint over streamable HTTP, authenticated with an API key. MCP is only a transport: every MCP tool call goes through the same policy, approval, verification and receipt path as the REST API. Connecting Voidex to external MCP servers as a client is planned.

Today Voidex ships a generic REST/HTTPS connector for APIs you control, signed outbound webhooks, a REST API with an OpenAPI 3.1 description, a TypeScript SDK and an MCP endpoint. The Acme Test CRM and Ledger are clearly labelled local simulators for development and benchmarking. Dedicated SaaS connectors are planned, not available yet.

Any agent that can call an HTTP API or an MCP server can use Voidex, regardless of model provider. Built-in model provider adapters are planned; the only provider shipped today is a deterministic demo used for testing. Voidex has no partnership with or endorsement from any model provider.

Yes, a TypeScript SDK with typed errors, safe retries for idempotent requests, run polling and webhook signature verification. A Python SDK is not available yet.

Credentials are encrypted with AES-256-GCM, bound to the owning organization, write-only through the API and never sent to the browser or to a model. They can be tested, rotated and revoked; revocation takes effect immediately for queued and new actions.

No. Voidex holds no compliance certification. It provides controls such as role-based access, approvals, append-only audit logs, receipts and retention settings that can support a customer’s own compliance work, and it publishes which controls are implemented and which are planned.

Encrypted action inputs and results are purged after the organization’s retention period, which owners configure. Receipts, which hold digests and outcomes rather than full payloads, are kept, and audit events are kept for the retention included in the plan.

There is a free Developer plan for sandbox use with simulated connectors. Paid plans start at US$29 per month for Builder and US$149 for Team, with Business and Enterprise priced with sales. Billing is currently in test mode, so no real charges can be made yet.

A verified action is an executed action whose receipt outcome is “verified”: every required check passed, including at least one deterministic or human-attested check. Blocked, rejected and failed actions are not verified actions, and retries of the same action count once.