concinno-skills-media

Media + smart-home agent skills for Concinno — OpenAI image generation, Replicate any-model dispatch, and Home Assistant state + service control via REST.

Status

MVP (0.1.0) — four tools, three upstream services. Fifth sub-package in the concinno-skills-* ecosystem (after -google, -chat, -content, -video), built to the same entry-points + Tool-protocol pattern.

Tools

Install

pip install concinno-skills-media

Deps pulled transitively: openai>=1.50, replicate>=1.0, httpx>=0.27, concinno>=2.15.1.

Credentials

All four tools read via concinno.core.credentials.CredentialStore (four-source precedence: runtime override > env var > ~/.concinno/credentials.json

None). Keys:

~/.concinno/credentials.json example

{
  "openai_api_key": "sk-...",
  "replicate_api_token": "r8_...",
  "homeassistant_url": "http://ha.local:8123",
  "homeassistant_token": "eyJhbGci..."
}

The HA token is a Long-Lived Access Token minted from your HA profile → Security page.

Cost warnings

ImageGenerate and ReplicateRun hit paid APIs. Every call logs a logger.warning line so CostTracker (if running in the host) can reconcile against billing:

• DALL-E 3: ~$0.04/image standard, ~$0.08 hd (1024² / 1792×1024)
• gpt-image-1: see OpenAI pricing
• Replicate: see the per-model pricing card on replicate.com

No local free-tier fallback is provided — use concinno-skills-content's KeywordExtract (local KeyBERT) if you want a free alternative for the text side.

Usage via Concinno ToolRegistry

When the consumer sets CONCINNO_LOAD_PLUGINS=1, the default registry auto-mounts all four tools:

import os
os.environ["CONCINNO_LOAD_PLUGINS"] = "1"

from concinno.tools.registry import get_default_registry

reg = get_default_registry()
names = set(reg.list_deferred())
assert {"ImageGenerate", "ReplicateRun",
        "HomeAssistantState", "HomeAssistantService"} <= names

img = reg.get("ImageGenerate")
out = img.call(
    action="create",
    prompt="a cozy cat on a windowsill, sunset",
    model="dall-e-3",
    size="1024x1024",
    quality="standard",
    n=1,
)
print(out["urls"][0])

Direct Python usage

from concinno_skills_media import (
    ImageGenerate, ReplicateRun,
    HomeAssistantState, HomeAssistantService,
)

# 1) Generate an image via DALL-E 3.
img = ImageGenerate().call(
    action="create",
    prompt="a 1990s anime cityscape at night",
    model="dall-e-3",
    size="1792x1024",
    quality="hd",
)
if img["ok"]:
    print(img["urls"][0])

# 2) Any Replicate model (Flux, SDXL, Wan, music, video, voice...)
flux = ReplicateRun().call(
    action="run",
    model="black-forest-labs/flux-schnell",
    input={"prompt": "an iguana on the beach, photorealistic"},
)
print(flux["output"])  # single URL or list of URLs depending on model

# 3) Home Assistant — read a light's state.
st = HomeAssistantState().call(
    action="get", entity_id="light.living_room"
)
print(st["state"], st["attributes"])

# 4) Home Assistant — turn that light on.
svc = HomeAssistantService().call(
    action="call",
    domain="light",
    service="turn_on",
    target={"entity_id": "light.living_room"},
    data={"brightness": 200, "color_name": "warm_white"},
)
print(svc["ok"], len(svc["changed"]))

All tools return either a structured {"ok": True, ...} dict or {"ok": False, "error": "..."} — never raise. Matches the rest of the Concinno tool ecosystem.

Safety notes

HomeAssistantService dangerous calls

Certain HA services mutate the installation itself or destroy history:

• homeassistant.restart — drops the HA process mid-work
• homeassistant.stop — shuts HA down entirely
• recorder.purge / recorder.purge_entities — drops historical data

The tool warns via logger.warning when these are invoked but does not block them — this is a power-user tool. If you want hard blocks, wrap in a custom guard / PreToolUse hook at the Concinno layer.

ImageGenerate safety rails

• n capped to 4 regardless of model (OpenAI itself caps DALL-E 3 at 1).
• size / quality validated against the published per-model enum; invalid values return {"ok": False, "error": ...} before any HTTP.
• Prompt capped at 4000 chars (DALL-E 3 limit — gpt-image-1 accepts much longer, but we keep the guard symmetric).

License

Apache-2.0. See LICENSE in the Concinno monorepo.