langchain-meragpt

LangChain integration for meraGPT Decider 1, a decision model. Ask a typed question about any text or JSON state (yes/no, pick one, or a score on a scale) and get back calibrated probabilities, not free text.

Decider 1 is state of the art among zero-shot models on the typed-decisions benchmark on Hugging Face. It costs $0.04 per 1M input tokens, and output is free.

This package gives you:

Installation

pip install langchain-meragpt
export MERAGPT_API_KEY=mg-live-...

Create a key at meragpt.com/dashboard/keys. New accounts get $1 of free credit.

Quickstart

from langchain_meragpt import MeragptDecider, choice, score, yes_no

decider = MeragptDecider()
answers = decider.invoke({
    "state": {"ticket": "I was charged twice for order A-1042. Fix it today or I'm cancelling."},
    "questions": {
        "topic": choice("What is the ticket mainly about?", {
            "billing": "Charges, refunds or invoices",
            "shipping": "Delivery of an order",
            "login": "Signing in or account access",
        }),
        "churn_risk": yes_no("Is the customer threatening to leave?"),
        "urgency": score("How urgent is this ticket?", ["low", "medium", "high"]),
    },
})

answers["topic"]["choice"]          # "billing"
answers["topic"]["probabilities"]   # {"billing": 0.96, "shipping": 0.035, "login": 0.004}
answers["churn_risk"]["noul"]       # probability of yes (0.5 here: a threat, but a conditional one)
answers["urgency"]["score"]         # expected level, 0 = low ... 2 = high (1.39)

Question types:

• yes_no(question): the answer is {"type": "noul", "noul": p_yes}.
• choice(question, options): up to 10 options, as a list of labels or a mapping of label to a one-line description (descriptions help). The answer has choice, confidence and probabilities.
• score(question, levels): an ordered scale, lowest first, up to 10 levels. The answer has score (the expected level), legend and probabilities.

Up to 64 questions can go in one call, and they are all answered against the same state. state can be a string or any JSON object.

Middleware

All middleware fails open: if a meraGPT call fails, the agent carries on as though the middleware weren't there, and a warning is logged. MeragptToolApprovalMiddleware can fail closed instead.

Triage and escalation (MeragptDecisionMiddleware)

from langchain.agents import create_agent
from langchain_meragpt import MeragptDecisionMiddleware, choice, yes_no

triage = MeragptDecisionMiddleware(questions={
    "topic": choice("What is the customer asking about?", {
        "billing": "Charges, refunds or invoices",
        "technical": "Bugs, errors or outages",
        "account": "Login, settings or cancellation",
    }),
    "escalate": yes_no("Is the customer asking for a human or threatening to leave?"),
})

agent = create_agent("openai:gpt-6-luna", tools=[...], middleware=[triage])
result = agent.invoke({"messages": [{"role": "user", "content": "..."}]})

if result["meragpt_decisions"]["escalate"]["noul"] > 0.7:
    hand_off_to_a_human(result)

Pass inject_into_prompt=True to add a one-line summary of the answers to the system message, so the agent can act on them too.

Approve or block tool calls (MeragptToolApprovalMiddleware)

from langchain_meragpt import MeragptToolApprovalMiddleware

guard = MeragptToolApprovalMiddleware(tools=["run_shell", "send_email"], threshold=0.5)
agent = create_agent("openai:gpt-6-luna", tools=[run_shell, send_email, search],
                     middleware=[guard])

Before a targeted tool runs, Decider sees the user's latest request, the tool's name and description, and the arguments, and answers "Does this tool call do what the user asked for?". Below the threshold, the tool doesn't run and the model gets an error ToolMessage explaining why. You can pass your own question, where "yes" approves the call. Keep it narrow, because one specific question separates far better than a composite "is this unsafe or unwanted?".

Route between models (MeragptRouterMiddleware)

from langchain_meragpt import MeragptRouterMiddleware

router = MeragptRouterMiddleware(routes={
    "simple": ("A short factual question or small talk", "openai:gpt-6-luna"),
    "complex": ("Multi-step reasoning, code or analysis", "anthropic:claude-opus-5-5"),
}, default="complex", min_confidence=0.6)

agent = create_agent("openai:gpt-6-luna", tools=[...], middleware=[router])

Stop a stuck agent (MeragptStopMiddleware)

from langchain_meragpt import MeragptStopMiddleware

agent = create_agent("openai:gpt-6-luna", tools=[...],
                     middleware=[MeragptStopMiddleware(min_tool_calls=4, threshold=0.7)])

After each model call that asks for more tool calls, the recent trace is checked: "Is the agent stuck repeating the same failing step?" and "Has the agent given up on the task?". If either question reaches the threshold, the run ends with a short "Stopped early" message instead of spending more calls. Pass your own questions to check something else.

As a tool

from langchain_meragpt import MeragptDecisionTool

agent = create_agent("openai:gpt-6-luna", tools=[MeragptDecisionTool()])

The agent calls meragpt_decide with state, question, kind (yes_no, choice or score) and options, and gets JSON back, for example {"p_yes": 0.83}.

Async

Every class has an async path (ainvoke, adecide, and the async middleware hooks).

What gets sent

Each decision sends the state and your questions to POST https://meragpt.com/v1/systemone. The middleware sends only what it needs:

• Decision and stop middleware: the last 20 turns.
• Approval middleware: the user's latest message, the tool call and the tool's description.
• Router middleware: the user's latest message.

Long texts are clipped. See the privacy policy for how requests are handled.

Configuration

Every class also takes api_key= and base_url=, or a shared client=MeragptClient(...). Requests refused with 429 or 503 while capacity scales up are retried, honouring Retry-After. Nothing is charged for a refused request.

Development

uv sync --all-groups
make test               # unit tests, network disabled
make integration_tests  # live API, needs MERAGPT_API_KEY
make lint

Links

• API reference: meragpt.com/docs/decider
• Model card, limits and evaluation: meragpt.com/models/decider-1
• Launch post: Introducing Decider 1