Piapi

PiAPI (Hermes — Creative)

PiAPI is a single-gateway inference platform: every media job is a task on POST /api/v1/task, polled via GET /api/v1/task/{task_id}. The LLM API at POST /v1/chat/completions is OpenAI-compatible and synchronous (with optional SSE streaming).

Prerequisites

export PIAPI_API_KEY="pk-..."
export PIAPI_WEBHOOK_SECRET="..."   # optional, webhook verification only
piapi-cli --help

There is no official PiAPI SDK — the CLI uses pure requests. The shim piapi-cli runs the bundled Python from ~/.local/share/piapi-skill/venv.

When to use

• A user prompt mentions Midjourney, Suno, Kling, Luma, Hailuo, Flux, Veo 3, Gemini Nano Banana, Seedance, Hunyuan, MMAudio, F5-TTS, Trellis, or Faceswap.
• The pipeline needs a vendor-agnostic media generator and PiAPI is the configured backend.
• Webhook callbacks from PiAPI need to be verified.
• An LLM call must go through PiAPI's OpenAI-compatible proxy.

When NOT to use

• The user already specified a competing platform (WaveSpeedAI, fal.ai, Replicate, Runway) — use that one instead.
• The job is a direct vendor SDK call (e.g., the user explicitly wants the official Suno API rather than PiAPI's wrapper).
• The flow is a creative-campaign chain that should stay in creative-campaign-media-studio; only invoke piapi for the leaf step that calls a model.

Quick reference

# Image — Midjourney
piapi-cli imagine "ultrarealistic mediterranean coast" --aspect-ratio 16:9 --process-mode fast --wait

# Image — Flux text-to-image
piapi-cli flux --prompt "cyberpunk alley at night" --model Qubico/flux1-schnell --wait

# Image — Gemini Nano Banana
piapi-cli submit --model gemini --task-type nano-banana-text-to-image \
  --input '{"prompt":"a calico cat in a cafe"}' --webhook-url ...

# Video — Kling
piapi-cli kling --image-url URL --prompt "slow dolly forward, cinematic" --duration 5 --wait

# Video — Veo 3
piapi-cli run --model veo3 --task-type txt2vid \
  --input '{"prompt":"a hummingbird in slow motion","aspect_ratio":"16:9","duration":8}'

# Video — Hunyuan with LoRA
piapi-cli run --model Qubico/hunyuan --task-type txt2video-lora \
  --input '{"prompt":"...","lora_url":"https://.../lora.safetensors"}'

# Music — Suno
piapi-cli suno --prompt "warm lo-fi instrumental, vinyl crackle" --wait

# Audio — F5-TTS
piapi-cli run --model Qubico/f5-tts --task-type txt2speech \
  --input '{"text":"Welcome to the demo.","voice_url":"https://.../voice.wav"}'

# 3D — Trellis (image to mesh)
piapi-cli run --model Qubico/trellis --task-type image-to-3d \
  --input '{"image_url":"https://.../object.png"}'

# Faceswap
piapi-cli faceswap --target-url T --swap-url F --wait

# Generic submit / wait / cancel / fetch
piapi-cli submit  --model <m> --task-type <t> --input '<json>' [--public|--private]
piapi-cli result  <task_id>
piapi-cli wait    <task_id> --timeout 1800 --poll-interval 2
piapi-cli cancel  <task_id>     # only valid while status is pending
piapi-cli run     --model <m> --task-type <t> --input '<json>'   # submit + wait

# LLM (OpenAI-compatible)
piapi-cli llm --model gpt-4o-mini \
  --message 'system:You are concise.' --message 'user:Hello.'

# Webhook verify (no HMAC — string compare)
piapi-cli verify-webhook --header-secret "$RAW_HEADER" --expected "$PIAPI_WEBHOOK_SECRET"

Python (no SDK — uses requests)

import os, requests, time

API = "https://api.piapi.ai"
H = {"x-api-key": os.environ["PIAPI_API_KEY"], "Content-Type": "application/json"}

def submit(model: str, task_type: str, payload: dict, *, mode="public", webhook=None):
    body = {"model": model, "task_type": task_type, "input": payload, "config": {"service_mode": mode}}
    if webhook:
        body["config"]["webhook_config"] = {"endpoint": webhook["url"], "secret": webhook["secret"]}
    r = requests.post(f"{API}/api/v1/task", json=body, headers=H, timeout=60)
    r.raise_for_status()
    return r.json()["data"]["task_id"]

def wait(task_id: str, timeout=1800, poll=2):
    deadline = time.time() + timeout
    while time.time() < deadline:
        r = requests.get(f"{API}/api/v1/task/{task_id}", headers=H, timeout=30)
        r.raise_for_status()
        data = r.json()["data"]
        status = (data.get("status") or "").lower()
        if status in {"completed", "complete", "success", "succeeded"}:
            return data
        if status in {"failed", "failure", "error", "canceled", "cancelled", "rejected"}:
            raise RuntimeError(f"task {task_id} failed: {data.get('error') or data}")
        time.sleep(poll)
    raise TimeoutError(task_id)

REST primitives

Auth header (lowercase canonical): x-api-key: <PIAPI_API_KEY>.

Status enum drift

PiAPI returns both capitalized and lowercase enums depending on the model family. Always lowercase before comparing.

• Midjourney / Kling / Faceswap: Pending | Processing | Staged | Completed | Failed
• Flux / Veo 3 / Gemini / Suno: pending | starting | processing | success | failed | retry

Terminal-success: completed, complete, success, succeeded. Terminal-failure: failed, failure, error, canceled, cancelled, rejected. Anything else means keep polling.

Service modes

• public — Pay-Per-Use (PPU). Default. Concurrency capped per pricing tier.
• private — Host-Your-Account (HYA). Required for Midjourney private mode (Discord token + channel ID + guild ID). Higher upfront commit, no per-image PPU charge after the seat fee.

Frequently-used model · task_type

Latency knobs

• Midjourney process_mode: fast (default), turbo (~half wait, costs more), relax (cheapest, queues for minutes).
• Flux: flux1-schnell is the speed model (4 steps); flux1-dev is the quality model (28 steps).
• Kling: mode = std | pro and duration = 5 | 10 seconds.
• Suno: set custom_mode=true to control lyrics + style; otherwise PiAPI fills both from the prompt.

Troubleshooting

• 401: PIAPI_API_KEY missing or wrong header casing — header must be exactly x-api-key.
• 402: out of credits or no active plan.
• 429: tier concurrency cap hit — back off, do not hammer.
• Status stuck Staged for Midjourney: that family stages an interactive grid; follow up with upscale / variation task_types.
• Webhook didn't fire: confirm endpoint returns 2xx quickly — PiAPI retries every 5s up to 3 attempts only.

References

• Overview: https://piapi.ai/docs/overview
• Per-family docs: https://piapi.ai/docs/{midjourney,flux,kling,suno,faceswap,veo3,gemini,...}
• LLM API: https://piapi.ai/docs/llm-api
• MCP server: https://piapi.ai/docs/mcp-server
• Pricing & rate limits: https://piapi.ai/pricing