FAQ

Straight answers, including to the questions whose honest answer is “not yet.” Status labels below match the legend in README.md.

​

Is the Execute API live today?

Yes, for invited partners and a deliberately narrow contract. Personal tenant keys can submit minimal durable task requests, read runs, resume SSE, request cancellation, read usage, configure automatic signed webhooks, and manage webhook deliveries. Turn mode, request-level overrides, organization execution, and mature receipts remain unavailable. Exact browser approval remains pending. Cancellation final states have passed separate production canaries. See status-and-roadmap.md.

​

What can I actually build on today, before the Execute API opens up?

Custom agents remain a separate available surface: create and publish an agent, then call that deployment with its deployment-scoped credential or supported embed channel. Its product-specific safety gates do not create generic public fabric tool/policy authority. See use-cases/agent-builders.md.

​

Can I manage agents, workflows, teams, and billing in the developer portal?

The portal-native personal agent/workflow routes, draft preview playground, profile/avatar/preferences, organization membership controls, subscription display, and eligible Stripe Checkout are source-ready. They are not documented as live until a compatible Worker deployment and authenticated production canaries are recorded. Organization API keys and execution remain unavailable. See Studio and developer console.

​

Can I upload a dataset and compare models?

Agent-scoped datasets are source-ready, pending their additive database migration, Worker deployment, and production isolation/immutability canary. Customer experiment execution and model comparison remain unavailable. A published dataset is immutable input evidence; it is not permission to call a model or spend credits. See Agent evaluation datasets.

​

Is the typed workflow DSL public?

No. The strict plan/compiler/interpreter foundation exists in source, but there is no public route, published package, deployed customer-definition authority, or connection to /v1/execute. The portal’s workflow page rehosts Praxa’s existing automation system; it is not the public DSL. See Typed workflow DSL.

​

How do approvals actually block an action?

Praxa’s existing product gates consequential actions fail closed. The first source-ready public form is narrower: an immutable browser action can project its exact instruction/steps, target host, and digest, then accept only the matching approve or deny decision. Generic tool approvals and the unified configurable policy engine are unavailable; public browser approval remains pending production activation. Details: approvals-and-policy/approvals.md.

​

What happens when a run hits its budget ceiling?

Funds are reserved before an admitted paid run executes. A ceiling or billing failure stops new work rather than becoming a surprise charge. Follow the run’s safe projection and usage readback for the public evidence available today; mature per-run receipts remain unavailable. See approvals-and-policy/budgets-and-limits.md.

​

Do you train on my data?

Praxa does not currently expose the planned cross-tenant learning system or its receipt consent labels as a completed fabric capability. The design keeps content, memory, and raw context out of cross-tenant signals and makes any future learning tenant-scoped or aggregated metadata, with explicit consent and evaluation gates. Details: security/data-handling.md.

​

Should I bring my own provider keys, or use bundled inference?

Bundled inference is the only active fabric funding path. BYO provider funding is planned but unavailable. Its intended security and fee model is documented in how-to/bring-your-own-keys.md.

​

What models do you route to?

The fabric is model-neutral by design. The current minimal task lane is pinned server-side; callers cannot yet submit a policy model-lane override. Mature receipts that record the selected model also remain unavailable.

​

Can I self-host Praxa?

No. The runtime is built on Cloudflare platform primitives and self-hosting is not offered. Dedicated account or region isolation, BYO provider keys, and full evidence export are planned enterprise capabilities, not current alternatives. See enterprise/deployment-and-isolation.md.

​

How do custom-agent keys relate to praxa_sk_ keys?

A custom-agent key is scoped to one existing deployment. praxa_sk_ is scoped to one admitted personal fabric tenant and exact endpoint scopes; the current execute body is a strict minimal task, not arbitrary agent/tool/policy control. Existing custom-agent keys keep working and neither credential gains the other’s authority. See keys-and-access/api-keys.md.

​

What’s the difference between mode: "turn" and mode: "task"?

The public contract defines only task, which returns a run_id and continues durably. Sending mode: "turn" returns 400 invalid_request, and /v1/chat is not routed. Candidate turn schemas in source are unserved. See api/execute.md.

​

What exactly is in a receipt?

That is the target receipt: what ran, policy version, exact approval evidence, and verification outcome. The current run and event resources are narrower safe projections; the mature receipt object is not shipped. See security/receipts-and-audit.md.

​

Is usage billed per token, per run, or per seat?

The active bundled lane reserves and settles credits from immutable execution economics. GET /v1/usage reports evidence-bounded run, token, call, and billable amount totals; it is not invoice authority. BYO fees are only a proposal because BYO is unavailable. See enterprise/billing-and-plans.md.

​

What happens if my system doesn’t respond to an approval in time?

After exact browser approval activation, a late, expired, stale, or already resolved decision is refused. The durable public contract does not promise a turn-style timeout event or a special expired run status; follow the run to its actual terminal projection. See approvals-and-policy/approvals.md.

​

Is the policy engine the same thing as approvals?

They are different planned layers. A unified engine would choose among auto, peer review, human approval, and deny; an approval mechanism would pause and resume the exact action. Today Praxa has multiple product-specific gates, while the public configurable engine and request policy override are unavailable. See approvals-and-policy/policy-engine.md.

​

Can I use this from Praxa Control, the desktop coding tool?

Not directly from the Control composer yet. A server-side script can use the partner-preview task API today; first-party Control cloud-lane integration remains planned. Local Codex or Claude Code sessions and their provider credentials remain separate. See use-cases/coding-and-desktop.md.

​

What happens if my connection drops mid-run?

Nothing is lost. Events are sequenced and replayable: reconnect to /v1/runs/:id/events with a Last-Event-ID header and you pick up exactly where you left off, with nothing re-delivered and nothing skipped. A task-mode run keeps going server-side regardless of whether anything is listening. See api/events.md.