Selora AI: Skills

Overview

A skill is an atomic capability the Selora AI agent can invoke. Skills are the agent’s imperative toolset: call an HTTP API, add a Home Assistant integration, create or modify an automation, read entity state, prompt the homeowner for input, escalate to an installer.

Skills are distinct from recipes . A recipe is a declarative package describing an outcome; a skill is one of the building blocks the pipeline uses to deliver that outcome. When the pipeline applies a recipe’s manifest and rendered package files, it dispatches each step to a skill.

Skills are inspired by OpenClaw skills and by the broader “tools” concept in agentic AI systems, narrowed to the smart-home domain.

Example

A recipe’s manifest.yaml declares:

Set up the nws integration with values resolved from the home’s latitude and longitude. Render package/automations/warning.yaml.j2 against the resolved roles and inputs, then commit the resulting automation to Home Assistant.

To apply that, the pipeline dispatches to several skills:

• ha.config.read_location() to fetch the homeowner’s coordinates.
• ha.integration.add(domain="nws", config={...}) to install the integration.
• ha.entity.wait_for_state(entity_id, state="<known>", timeout=60) to verify the integration came online.
• ha.automation.create(name=..., trigger=..., action=...) to register the automation.

Each skill is a typed, tested, sandboxed function. The pipeline picks them; the recipe author never names them.

Customer value

• Reliability: skills are typed and tested, so the agent’s actions are predictable and auditable. Failure modes are explicit.
• Security: skills sit at well-defined trust boundaries. Sensitive skills (locks, payments, network config) require explicit user confirmation. Other skills are silent.
• Local-first: skills run on the homeowner’s Selora Hub. No cloud call-out is required for the core skill set.
• Extensible: third parties can contribute skills the same way they contribute integrations. New domains (e.g. an irrigation skill set) expand the agent’s reach without changing recipe authoring.

Scope (first iterations)

• Core skill catalog: the minimum set required to install and uninstall the marquee recipes: HA integration config, automation create/delete, entity state read/write, HTTP fetch, user prompt, recipe state read/write.
• Skill specification: type signature, side-effect classification (read-only, mutates HA state, mutates external service, prompts user), rollback semantics.
• Trust tiers: skills are tagged by sensitivity. The agent confirms with the homeowner before invoking a sensitive skill (anything that touches locks, alarms, payments, or external authentication).
• Skill discovery: the agent advertises its installed skill catalog so recipe reviewers can verify that a recipe’s directives are expressible in terms of supported skills.
• Local testing harness: a way to run a recipe end-to-end against a mocked Home Assistant to verify that each manifest step and rendered package file resolves to a real skill invocation and that the recorded artifacts can be cleanly reversed on uninstall.

Architecture (directional)

• Skills are implemented as code inside the embedded Claw on each Selora instance. They are not user-authorable from the marketing site or from Connect.
• Each skill is a function with a typed input schema, a typed output schema, an idempotency declaration, and an optional cleanup procedure the agent calls when uninstalling.
• The pipeline’s runtime lists the installed skills, their type signatures, and short descriptions. For each manifest step and rendered package artifact, it selects the matching skill(s) and supplies the required parameters.
• Skill updates ship through the same channel as Selora AI itself (Hub firmware updates or HACS package updates), not through the recipe library.

Target customers

• Homeowners: indirectly. The skill catalog determines what the agent can do, but homeowners interact only with recipes and chat.
• Recipe authors: knowing the available skills helps authors write realistic manifests and package templates the pipeline can apply.
• Third-party skill contributors: developers who want to expand Selora AI into new domains (irrigation, electric vehicles, accessibility) by shipping additional skills.

Open questions and risks

• Skill versioning: recipes declare a minimum agent version they require. How granular should this be (agent version vs individual skill versions)?
• Sensitive-skill UX: when the agent calls a skill that needs explicit confirmation, how is the prompt phrased so the homeowner understands the action without being asked to read code?
• Sandboxing: third-party skills introduce code execution risk. How to enforce skill boundaries (network egress, filesystem access, HA API scope) at runtime.
• Cross-skill state: how skills coordinate when multiple recipes touch the same entities (e.g., two recipes both wanting to set up a siren).
• Discoverability: when a recipe references a skill that is missing on the homeowner’s instance, the agent should refuse to install the recipe with a clear explanation, not fail mid-way.

Related

• Selora AI: Library of Recipes (Epic #52): Roadmap — the declarative layer that drives skill invocations.
• Selora AI: Embedded Claw (Epic #45): Roadmap — the runtime where skills live and execute.
• Selora AI: Create Scenes (Epic #46): Roadmap — an early skill that bundles scene-creation primitives.
• Selora AI: Suggest Integrations & Devices (Epic #51): GitLab epic #51 — relies on skills to enumerate the homeowner’s setup and recommend hardware.