Skip to main content
SMEme
Sign in Try free

For builders

Add a reasoning step to your AI workflows.

Call SMEme over MCP to evaluate approved decision procedures inside your own harness. Your agent gathers the facts. SMEme returns what follows, what’s unresolved, and the trace behind it.

Read the MCP docs
Works with LangGraph n8n Any MCP client
your harnessabridged
smeme_reasoning_capabilities() smeme_reasoning_guidance_get() smeme_reasoning_list() smeme_reasoning_evaluate(decision_tree_id) while harness_next == "continue_evaluate": smeme_reasoning_evaluate_continue(…) → report
Parameters and response shapes: see the MCP tool reference.

Five steps, one loop.

The language model proposes. A rule or a person admits. SMEme evaluates. Missing facts send the loop back for more.

  1. 1 · bootstrapGet readyCheck capabilities, cache the calling guidance, list deployed trees.
  2. 2 · proposeAgent proposes factsFrom documents, tools, and conversation.
  3. 3 · admitFacts are admittedBy rule or by a human reviewer. Nothing is evaluated until it’s admitted.
  4. 4 · evaluateSMEme reasonsAdmitted facts against the approved tree. Deterministic.
  5. 5 · reportRoute the resultConclusions, open questions, and trace go downstream.
Facts still missing? Report loops back to propose.
Isolated evaluation required When a step needs independent verification, SMEme stops instead of faking it. Your harness gets an explicit stop, never a guessed answer.

Here is one LangGraph example of that harness. The code and how to run it are in the LangGraph docs.

Demo coming soon
SMEme as a workflow step

See it run in a pipeline

A new intake form triggers the workflow. The agent proposes facts from the form and email, a reviewer admits them, and SMEme evaluates: Accept with escalation, plus the trace. The workflow routes on the result.

The same step, written in LangGraph, is in the LangGraph technical preview.

More than a verdict.

Because the tree is compiled to constraints, your harness can ask follow-up questions programmatically.

smeme_reasoning_list_conclusionsWhat follows, what’s ruled out, what’s still open. smeme_reasoning_decisive_supportWhich facts determined the result. smeme_reasoning_what_ifChange a fact and compare the outcome. smeme_reasoning_how_to_reachWhat would need to change for another outcome. smeme_reasoning_edit_affects_pathWhether a proposed edit would change this case’s path. template_get → validate_answers → evaluate_answersBulk and audit runs over a worksheet of cases.

Guarantees

Deterministic

Same admitted facts and approved tree, same result. Testable like any other step.

Logic outside the prompt

Trees are versioned and approved apart from the model. Swap models without rewriting rules.

Human in the loop

The admit step is where a reviewer signs off on the facts.

Audit-ready trace

Every report carries the reasoning behind it.

HostedConnect in minutes

One MCP endpoint with OAuth. Free tier with 3 decision trees and 150 MCP calls.

MCP docs

Put one approved procedure in your pipeline.

Create a free account Read the docs