Blueprints

Modules and the SDK stay vendor-neutral by design. Every vendor concern lives in blueprints/. Each blueprint takes a ComposedPrompt from @lokomotif/sdk and produces the format its target runtime expects, with the composition hash preserved end-to-end so traces correlate back to the originating Lokomotif composition.

The four shipped runtimes

Blueprint	Output	Vendor dep
`@lokomotif/blueprint-anthropic-sdk`	Anthropic Messages API call	`@anthropic-ai/sdk` (peer)
`@lokomotif/blueprint-dify`	Workflow YAML (Dify import-ready)	none — structural transform
`@lokomotif/blueprint-n8n`	Workflow JSON (n8n import-ready)	none — structural transform
`@lokomotif/blueprint-langgraph`	LangGraph `StateGraph`	`@langchain/langgraph` (peer)

Anthropic SDK

import Anthropic from '@anthropic-ai/sdk';
import { compose, loadModules } from '@lokomotif/sdk';
import { runWithAnthropic } from '@lokomotif/blueprint-anthropic-sdk';
 
const composed = compose(loadModules([...], { modulesDir }));
 
const message = await runWithAnthropic(composed, 'Vakayı özetle.', {
  client: new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY }),
  model: 'claude-sonnet-4-6',
});

The composed prompt becomes the Anthropic Messages system field; the user input becomes a single user message. Multi-turn conversation remains the caller’s responsibility. The composition hash is exposed so emitters can record lokomotif.flow.composition_hash (see @lokomotif/otel-schema).

Dify

import { compose } from '@lokomotif/sdk';
import { adaptToDify, renderDifyYaml } from '@lokomotif/blueprint-dify';
import { writeFileSync } from 'node:fs';
 
const composed = compose(modules);
const definition = adaptToDify(composed, { appName: 'aml-review' });
writeFileSync('aml-review.dify.yaml', renderDifyYaml(definition));

The output is a workflow-mode Dify definition with start → llm → end nodes wired in series. Import via the Dify dashboard (Apps → Create from DSL).

n8n

import { compose } from '@lokomotif/sdk';
import { adaptToN8n, renderN8nJson } from '@lokomotif/blueprint-n8n';
 
const composed = compose(modules);
const workflow = adaptToN8n(composed, { workflowName: 'aml-review' });
writeFileSync('aml-review.n8n.json', renderN8nJson(workflow));

The output is an n8n workflow with a manual trigger → Anthropic chat node → set output node, all wired in series. Import via Workflows → Import from File.

LangGraph

import { compose } from '@lokomotif/sdk';
import { buildStateGraph } from '@lokomotif/blueprint-langgraph';
 
const composed = compose(modules);
const graph = buildStateGraph(composed, {
  llm: async (state) => callYourLLM(state.system_prompt, state.user_input),
});
 
const compiled = graph.compile();
const result = await compiled.invoke({ user_input: 'analyze this' });

The graph topology is fixed: compose → execute → audit. The compose node injects the system prompt into state. The execute node calls the caller-supplied llm callback. The audit node records the composition hash and module manifest for downstream observability.

Adding a new runtime

A new blueprint requires an RFC. Why: the runtime list is the surface that operators evaluate the Kit against; adding entries casually dilutes the signal. The RFC documents the runtime’s adapter contract, the test strategy without external API access, and the maintenance burden.