Question 1

Is my data really private?

Accepted Answer

Yes. The engine is stateless - it reads a soul_state in and returns a fresh one out, writing nothing to disk. On the hosted API we store your account, aggregate usage counts, and billing state, but never your messages or your soul_state. With the SDK or self-hosted deployment, your data never touches us at all.

Question 2

What exactly is soul_state, and why do I store it?

Accepted Answer

A compact JSON blob (under 48KB) holding calibration histograms, Bayesian beliefs, the user model, the curiosity ledger, and the neural meta-controller weights. You persist it and send it back each turn - that round-trip is the learning. Because it lives with you, a competitor can't lift it, and if you leave you keep everything you learned.

Question 3

Which model providers work?

Accepted Answer

You bring your own: OpenAI, Anthropic, or a local model, all behind one adapter interface. LEO Soul passes your key through for that request and never stores it.

Question 4

Will it slow down my agent?

Accepted Answer

Calm turns take a fast path - roughly one model call, near pass-through latency. Heavy scrutiny is spent only when triage detects real stakes or anomalies, and the trade-off self-tunes per deployment.

Question 5

What happens if LEO Soul itself goes down?

Accepted Answer

Your SDK decides, and you set it explicitly. on_unavailable defaults to fail_closed - the call raises, so nothing unverified reaches a user. Set it to fail_open and the client calls your model provider directly and returns the answer flagged degraded, so availability wins but you always know the answer was not checked. This has to live in your code rather than our dashboard: during our outage we cannot serve you a setting. We also notify affected accounts by email, an in-app banner and a service.degraded webhook, and an external heartbeat independent of our own infrastructure detects a total outage.

Question 6

Will I be emailed if my API calls start failing?

Accepted Answer

Yes. If you run out of turns you get a warning email at 80% and 100%, then a distinct notice the moment calls actually start being refused, plus up to two follow-ups a day apart while you are still blocked and a usage.blocked webhook. For incidents and planned maintenance we email you, raise an in-app banner for every member of your workspace, and fire a service.degraded webhook. Add an operations or on-call address in Settings and all of it goes to your rota as well as to you.

Question 7

How do I tell your outage apart from my model provider failing?

Accepted Answer

By the status code. A failure in your own model provider - an expired key, their rate limit, their outage - comes back as 502 with an upstream block naming the provider, the kind of failure and whether retrying helps. A 500 means the fault is ours. They used to look identical, which sent people debugging the wrong system.

Question 8

What happens if a module errors?

Accepted Answer

Every module runs inside a fail-open guard. On any error it degrades to a safe pass-through and records itself as degraded in the trace, so one failed check never costs you the turn. Regulated deployments can switch this to fail-closed. That covers a module failing; if LEO Soul itself is ever unreachable, your SDK's on_unavailable setting decides what your app does - it defaults to failing closed so nothing unchecked reaches a user.

Question 9

What happens when I hit my monthly turn limit?

Accepted Answer

Your dashboard shows a usage bar and flags you at 80%, so you see it coming. On a paid plan we allow a little headroom above the quota before blocking, so a production integration isn't cut off mid-request; on the Free plan the allowance is a hard limit. Upgrade whenever you're ready - Enterprise is unlimited.

Question 10

How do API keys and agents work?

Accepted Answer

One API key powers one agent, each with its own memory and analytics. Free includes a single key; Pro includes 10 and Scale 50. You can rotate a key in one click if it leaks, which issues a fresh secret and disables the old one.

Question 11

Can I self-host it?

Accepted Answer

Yes. The core engine is pure-Python with zero runtime dependencies, so it already runs entirely on your side via the SDK. Enterprise self-hosted ships the full platform as a Docker image for VPC / on-prem, with SSO and a data-processing agreement. Prefer not to run anything? Enterprise Hosted gives you the same platform fully managed on our cloud - you just log in.

Question 12

Do you host Enterprise for us, or do we run it?

Accepted Answer

Your choice. Enterprise Hosted is fully managed on our cloud - nothing to deploy, you just sign in. Enterprise Self-hosted runs the Docker image inside your own network for strict data residency or air-gap. Both include the same unlimited usage, SSO/SAML + SCIM, DPA/BAA, dedicated engineer and SLA, at custom pricing.

Question 13

How is this different from a 'be careful' prompt?

Accepted Answer

Prompts are vibes. LEO Soul runs real algorithms: semantic entropy over meanings, Bayesian beliefs, Brier-calibrated confidence, a conformal error bound, and an online neural model that learns which situations fail in your deployment specifically.

Question 14

How does LEO Soul compare to guardrails or eval tools like Guardrails AI, Galileo or Patronus?

Accepted Answer

Those tools mostly score or block outputs after the fact. LEO Soul is a metacognitive layer that runs before the answer is committed: it estimates its own uncertainty, decides whether to answer / ask / confirm / refuse, and carries a portable memory that learns your domain over time. You can run it alongside an eval or observability stack - it's the decision layer, not just a filter.

Question 15

What do I get on the Free plan?

Accepted Answer

A real, permanent plan: one agent, 100 turns a month, the full decision loop, and the in-browser Playground. Free also currently includes a bonus of paid engine features - calibrated confidence, reliability reports, and cost analytics - so you can feel the engine before paying. Pro adds self-learning, tool-call guardrails, and much higher limits.

Question 16

Do I need to change my model or rewrite my prompts?

Accepted Answer

No. You keep your own model and prompts; LEO Soul wraps the turn and decides how to handle it. With the drop-in gateway you just change your base URL. There is no LEO Soul model to choose - you bring your own.

Questions, answered