meshapi

Mesh API integration

fiebatt can route its text planning and agent tool-calling layer through Mesh API, while keeping the rest of the video pipeline unchanged.

Mesh API is OpenAI-compatible, so the integration is intentionally small:

from openai import AsyncOpenAI

client = AsyncOpenAI(
    api_key=os.environ["MESH_API_KEY"],
    base_url=os.environ.get("MESH_API_BASE_URL", "https://api.meshapi.ai/v1"),
)

Default model

The project defaults to:

deepseek/deepseek-v3.2

That model is used for the agent's text reasoning, prompt rewriting, tool selection, and structured edit planning when MESH_API_KEY is present.

Vision-specific calls still prefer the existing vision provider when a request contains image payloads, because the default Mesh model above is a text model.

Environment

MESH_API_KEY="rsk_..."
MESH_API_BASE_URL="https://api.meshapi.ai/v1"
MESH_MODEL="deepseek/deepseek-v3.2"
MESH_VIDEO_MODEL="google/veo-3"
MESH_VIDEO_ENDPOINT="/video/generations"
USE_AI_STUBS="false"

With those values set, /api/agent/chat will use Mesh API for the agent loop.

To route video generation through Mesh as well:

VIDEO_GEN_PROVIDER="meshapi_veo"

The adapter sends Mesh's documented content array, polls the returned task at /v1/video/generations/{id}, and downloads content.video_url after the task succeeds. The model remains configurable because Mesh model availability can vary by account. Keep MESH_VIDEO_MODEL aligned with the Veo model ID shown in your Mesh dashboard.