muapi-langchain 0.1.0

LangChain integration for muapi.ai — generative media tools, document loader, and Deep Agents recipe

muapi-langchain

LangChain integration for muapi.ai — 250+ generative media models behind a single key, exposed as LangChain tools, a document loader, and a Deep Agents recipe.

Install

pip install muapi-langchain

# For the Deep Agents demo:
pip install "muapi-langchain[deepagents]"

Or install directly from the repo (latest main):

pip install "git+https://github.com/SamurAIGPT/muapi-cli.git#subdirectory=integrations/langchain"

Set your API key:

export MUAPI_API_KEY="..."     # or: muapi auth configure

What's in the box

4 tools (capability gradient)

Tool	What it does	Costs credits?
muapi_select	Rank models + skills for an intent	No
muapi_generate	Single-shot generation (image / video / audio / edit)	Yes (per call)
muapi_run_skill	Named multi-step recipe (UGC ad, storyboard, …)	Yes (recipe)
muapi_creative_agent	Open-ended brief → planner → executor	Yes (variable)

from muapi_langchain import muapi_select, muapi_generate

print(muapi_select.invoke({
    "intent": "cinematic product photo of a sneaker",
    "kind": "image", "tier": "best", "limit": 3,
}))

print(muapi_generate.invoke({
    "prompt": "A glossy sneaker on a wet street at neon-lit night",
    "kind": "image", "tier": "best",
}))

MuapiAssetLoader — document loader

Hydrate prior muapi generations as Documents for RAG / eval.

from muapi_langchain import MuapiAssetLoader

docs = MuapiAssetLoader(request_ids=["req_abc", "req_def"]).load()
for d in docs:
    print(d.metadata["url"], "·", d.page_content[:80])

MuapiCostCallback — budget tracking

Pipe every muapi tool call through a callback that accumulates credit spend and aborts the agent when a budget cap is hit.

from muapi_langchain import MuapiCostCallback

cost_cb = MuapiCostCallback(
    budget_credits=500,
    on_event=lambda evt, payload: print(evt, payload),
)
agent.invoke({...}, config={"callbacks": [cost_cb]})
print(cost_cb.summary())

Deep Agents recipe

The recommended pattern: cheap tools (select, generate) on the main planner, heavy tools (run_skill, creative_agent) on a creative-specialist subagent, with interrupt_on for human approval of open-ended creative work.

See examples/deep_agents_demo.py for the full runnable example.

from deepagents import create_deep_agent
from muapi_langchain import PLANNER_TOOLS, SPECIALIST_TOOLS

agent = create_deep_agent(
    model=...,
    tools=PLANNER_TOOLS,
    subagents=[{
        "name": "creative-specialist",
        "description": "Heavy muapi workflows and open-ended briefs.",
        "system_prompt": "...",
        "tools": SPECIALIST_TOOLS,
    }],
    interrupt_on={
        "muapi_creative_agent": {"allowed_decisions": ["approve", "edit", "reject"]},
    },
)

Decision tree

User brief
  ├─ Don't know which model/skill?  → muapi_select          (free)
  ├─ Single asset, clear prompt?    → muapi_generate
  ├─ Matches a known recipe?        → muapi_run_skill
  └─ Multi-asset / multi-modal?     → muapi_creative_agent  (interrupt_on)

License

MIT.