genai-calling 0.1.6

Single-endpoint GenAI SDK (multi-provider, multimodal)

genai-calling

中文文档：README_ZH.md

One interface for calling multimodal models; four ways to use: Skill, MCP, CLI, SDK.

Features

• Multi-provider: OpenAI, Google (Gemini), Anthropic (Claude), Aliyun (DashScope/Bailian), Volcengine (Doubao/Ark), Tuzi
• Multimodal: text/image/audio/video input and output (model-dependent)
• Unified API: a single Client.generate() for all providers
• Streaming: generate_stream() for incremental output
• Tool calling: function tools (model/provider-dependent)
• JSON Schema output: structured output (model/provider-dependent)
• MCP Server: Streamable HTTP and SSE transport
• Security: SSRF protection, DNS pinning, download limits, Bearer token auth (MCP)

Installation

pip install genai-calling

Python import package is gravtice.

For development:

pip install -e .
# or (recommended)
uv sync --group dev

Skill (External Repository)

The standalone skill is no longer bundled in this repository.

Preferred install name:

npx skills add gravtice/nous-skills -s genai-calling

Legacy catalogs may still expose the old entry name:

npx skills add gravtice/nous-skills -s nous-genai

Skill repository: https://github.com/gravtice/nous-skills

Configuration (Env Vars, Zero-parameter)

Configuration is managed via environment variables.

You can set env vars in two ways:

1. Runtime env vars (inline or exported in shell)
2. Env files (.env.local, .env.production, .env.development, .env.test) and the global fallback ~/.genai-calling/.env

Runtime example (inline):

GENAI_CALLING_OPENAI_API_KEY=... uv run genai --model openai:gpt-4o-mini --prompt "Hello"

When env files are used, SDK/CLI/MCP loads them automatically with priority (high -> low):

.env.local > .env.production > .env.development > .env.test > ~/.genai-calling/.env

Process env vars override both project and global env files (the loader uses os.environ.setdefault()).

Use ~/.genai-calling/.env for user-wide shared defaults such as API keys. Keep worktree-specific settings such as ports in project-local .env.local.

Minimal .env.local (OpenAI only):

GENAI_CALLING_OPENAI_API_KEY=...
GENAI_CALLING_TIMEOUT_MS=120000

See docs/CONFIGURATION.md for all options, or copy .env.example to .env.local.

Quickstart

CLI (fastest, unified API, agent-friendly)

# List available models by capabilities (out=text/image/audio/video/embedding)
uv run genai model available --all

# Text generation
uv run genai --model openai:gpt-4o-mini --prompt "Hello"

# Image understanding (image -> text)
uv run genai --model openai:gpt-4o-mini --prompt "Describe this image" --image-path ./examples/demo_image.png

# Image generation (text -> image file)
uv run genai --model openai:gpt-image-1 --prompt "A red cube on white background, minimal" --output-path ./out.png

# Speech-to-text (audio -> text)
uv run genai --model openai:whisper-1 --audio-path ./examples/demo_tts.mp3

# Text-to-speech (text -> audio file)
uv run genai --model openai:tts-1 --prompt "Hello from genai-calling" --output-path ./out.mp3

# Video generation (text -> video; async style)
uv run genai --model openai:sora-2 --prompt "A paper boat sailing on a rain puddle, cinematic" --no-wait
# ...later
uv run genai --model openai:sora-2 --job-id "<job_id>" --output-path ./out.mp4 --timeout-ms 600000

SDK: Text generation

from gravtice import Client, GenerateRequest, Message, OutputSpec, Part

client = Client()
resp = client.generate(
    GenerateRequest(
        model="openai:gpt-4o-mini",
        input=[Message(role="user", content=[Part.from_text("Hello!")])],
        output=OutputSpec(modalities=["text"]),
    )
)
print(resp.output[0].content[0].text)

SDK: Streaming

import sys
from gravtice import Client, GenerateRequest, Message, OutputSpec, Part

client = Client()
req = GenerateRequest(
    model="openai:gpt-4o-mini",
    input=[Message(role="user", content=[Part.from_text("Tell me a joke")])],
    output=OutputSpec(modalities=["text"]),
)
for ev in client.generate_stream(req):
    if ev.type == "output.text.delta":
        sys.stdout.write(str(ev.data.get("delta", "")))
        sys.stdout.flush()
print()

SDK: Image understanding

from gravtice import Client, GenerateRequest, Message, OutputSpec, Part, PartSourcePath
from gravtice import detect_mime_type

path = "./cat.png"
mime = detect_mime_type(path) or "application/octet-stream"

client = Client()
resp = client.generate(
    GenerateRequest(
        model="openai:gpt-4o-mini",
        input=[
            Message(
                role="user",
                content=[
                    Part.from_text("Describe this image"),
                    Part(type="image", mime_type=mime, source=PartSourcePath(path=path)),
                ],
            )
        ],
        output=OutputSpec(modalities=["text"]),
    )
)
print(resp.output[0].content[0].text)

SDK: List available models

from gravtice import Client

client = Client()
print(client.list_all_available_models())

Providers

Provider	Notes
openai	GPT-4, DALL·E, Whisper, TTS
google	Gemini, Imagen, Veo
anthropic	Claude
aliyun	DashScope / Bailian (OpenAI-compatible + AIGC)
volcengine	Ark / Doubao (OpenAI-compatible)
tuzi-web / tuzi-openai / tuzi-google / tuzi-anthropic	Tuzi adapters

Binary output

Binary Part.source is a tagged union:

• Input: bytes/path/base64/url/ref (MCP forbids bytes/path)
• Output: url/base64/ref (SDK does not auto-download to disk)

If you need to write to file, see examples/demo.py (_write_binary()), or reuse Client.download_to_file() for the built-in safe downloader.

CLI & MCP Server

# CLI
uv run genai --model openai:gpt-4o-mini --prompt "Hello"
uv run genai model available --all

# Tuzi Chirp music
uv run genai --model tuzi-web:chirp-v3-5 --prompt "Lo-fi hiphop beat, 30s" --no-wait
# ...later
uv run genai --model tuzi-web:chirp-v3-5 --job-id "<job_id>" --output-path demo_suno.mp3 --timeout-ms 600000

# MCP Server
uv run genai-mcp-server                    # Streamable HTTP: /mcp, SSE: /sse
uv run genai-mcp-cli tools                 # Debug CLI

Security

• SSRF protection: rejects private/loopback URLs by default (GENAI_CALLING_ALLOW_PRIVATE_URLS=1 to allow)
• DNS pinning: mitigates DNS rebinding
• Download limit: 128MiB per URL by default (GENAI_CALLING_URL_DOWNLOAD_MAX_BYTES)
• Bearer token auth: for MCP server
• Token rules: fine-grained access control

Testing

uv run python -m pytest tests/ -v

Docs

• Configuration
• Contributing
• Changelog

License

MIT