Skip to main content

​
When to use it

  • Short user inputs need to be mapped to a handful of flows before you invest in full orchestration.
  • You want to gate automation on a confidence score (only auto-run when the intent is clear, otherwise escalate).
  • You need structured metadata—like extracted entities or a human-readable reason—to feed into downstream logic.
  • You want deterministic categorisation (embeddings) or richer explanations (LLM) without building a bespoke classifier.

​
Defining intents

Every classifier consumes a list of Intent objects: 
from mcp_agent.workflows.intent_classifier.intent_classifier_base import Intent

INTENTS = [
    Intent(
        name="fetch_file",
        description="Retrieve the contents of a file from the filesystem MCP server.",
        examples=[
            "show me README.md",
            "open src/app.py",
            "cat /var/log/system.log",
        ],
        metadata={"priority": "high", "team": "infra"},
    ),
    Intent(
        name="general_question",
        description="Answer an informational question without tool use.",
        examples=["what is MCP?", "explain the router pattern"],
    ),
]
  • description gives the classifier context and is surfaced in tracing metadata.
  • examples dramatically improve accuracy—provide several phrasing variants.
  • metadata is propagated to the result so you can attach business logic (e.g. SLA, handoff target).

​
Choosing a classifier

VariantFactory helperBest forOutput extras
LLM-basedcreate_intent_classifier_llm(...)Highest quality natural language understanding, explanations, entity extractionconfidence (low/medium/high), p_score, reasoning, extracted_entities
Embedding-basedcreate_intent_classifier_embedding(...)Deterministic scoring, lower latency, custom embedding providersp_score (0–1 similarity)
LLM classification enforces a strict JSON schema (StructuredIntentResponse), ensuring stable output even under temperature.

​
Quick start

from mcp_agent.app import MCPApp
from mcp_agent.workflows.factory import (
    create_intent_classifier_embedding,
    create_intent_classifier_llm,
)
from mcp_agent.workflows.intent_classifier.intent_classifier_base import Intent

app = MCPApp(name="intent_demo")
INTENTS = [...]  # see definition above

async def main():
    async with app.run() as running_app:
        llm_classifier = await create_intent_classifier_llm(
            intents=INTENTS,
            provider="openai",
            classification_instruction="Return at most one intent unless the user explicitly asks for multiple.",
            context=running_app.context,
        )

        embedding_classifier = await create_intent_classifier_embedding(
            intents=INTENTS,
            provider="openai",  # or "cohere"
            context=running_app.context,
        )

        request = "Could you open README.md for me?"
        llm_result = (await llm_classifier.classify(request, top_k=2))[0]
        emb_result = (await embedding_classifier.classify(request, top_k=2))[0]

        return {
            "llm_intent": llm_result.intent,
            "llm_confidence": llm_result.confidence,
            "llm_reasoning": llm_result.reasoning,
            "embedding_intent": emb_result.intent,
            "embedding_score": emb_result.p_score,
        }

​
Working with results

  • LLM classifier returns LLMIntentClassificationResult with:
    • intent: matched intent name.
    • confidence: "low", "medium", "high" (auto-quantised from raw scores).
    • p_score: continuous probability (0–1).
    • reasoning: short explanation.
    • extracted_entities: optional name/value pairs surfaced by the LLM.
  • Embedding classifier returns IntentClassificationResult with intent and p_score. Sort or threshold the score to decide automation boundaries.
Both variants support top_k, letting you offer alternatives to a human or feed multiple candidates into a downstream router.

​
Integrating with the router

Intent classifiers and routers pair naturally: classify first, then route using a richer skill set. 
intent = (await llm_classifier.classify(request, top_k=1))[0]
if intent.confidence != "high":
    return "Escalating to human – intent unclear."

decisions = await router.route(
    f"[intent={intent.intent}] {request}",
    top_k=3,
)
The intent name/metadata can be prepended to the router prompt (as above) or used to select different router instances entirely.

​
Tuning and operations

  • Override classification_instruction to bias LLM behaviour (hierarchical intents, abstain thresholds, multilingual hints).
  • Pass request_params=RequestParams(strict=True, temperature=0) to disable sampling variance for high-stakes automation.
  • Pre-compute embeddings for cold start by calling await classifier.initialize() at app startup.
  • Record tracing output (otel.enabled: true) to inspect intent descriptions, examples, and resulting confidence scores per request.

​
Example projects

⌘I