Skip to content

What AsteronIris is not

Listing non-goals is more useful than listing capabilities. Capabilities drift; non-goals define the project’s spine.

AsteronIris never denies being an AI. Prompts, persona definitions, and the pre-send verifier all enforce this. A companion whose first move is to claim humanity cannot build a trust relationship on anything real; the project refuses that trade.

Being openly AI is the enabling constraint, not a limitation. Everything else — memory, affect, persona continuity — becomes meaningful because the user is not being tricked about what they are talking to.

AsteronIris is a single-operator runtime. One person, or a small trusted team, runs the daemon on a machine they control. The threat model assumes the operator is trusted. The security boundaries protect:

  • The LLM and its tools from doing things outside the operator’s consent
  • External ingress (webhooks, A2A) from forging trust signals the edge did not verify
  • Secrets from leaking through outbound surfaces

That is a different shape than “a cloud product serving arbitrary tenants”. The runtime has tenant-scoped admin and operator concepts so local workspaces and bindings stay separated, but those are not a public SaaS isolation model. Arbitrary-tenant hosting is not supported and is not a near-term direction.

The runtime has tools, a tool registry, subagents, and a governed tool loop. None of that is the product. They exist because a companion that cannot look anything up or take small grounded actions drifts into vague small talk; they are continuity infrastructure, not agentic showpieces.

If the question is “how many autonomous steps can this plan and execute”, AsteronIris is not the answer. If the question is “how much grounding can this runtime hold across a month of conversation without losing who it is”, it is.

The system does not expose a user-editable system prompt as the main configuration path. Character behavior is shaped through:

  • Persona definitions (durable identity)
  • Affect topology (how emotions route through character-specific geography)
  • Memory ingestion and recall policy
  • Pre-send verifier rules
  • Tool allowlists and policy

Prompt text is generated from those. Tweaking prompt text directly is possible but is treated as a short-term lever, not a design interface.

Voice adapters exist. They are not where the project is proven. Text is the primary modality because:

  • Memory and continuity are easier to verify in text
  • The companion’s character has to read right in the most scrutinized channel first
  • Voice layers cleanly on top of a working text runtime; the inverse is not true

Voice-first competitors are solving a different product.

The desktop/ console exists as an operator surface — governance, diagnostics, admin workflows, memory review. It is not where users meet the companion. Users meet the companion in Discord (primary), or in other channels that feed the same turn pipeline.

Active development. APIs and behavior can change between commits. Treat the README’s status matrix (Stable / Beta / Alpha) as current truth; this site is intentionally conceptual and does not track minor version shifts.