Call an LLM

Use this when a step needs open-ended text: summarise, draft, extract, plan. For small choices (which branch, is this safe) a decision plugin is cheaper. Prerequisites: a server on http://localhost:9858 whose build includes the LLM_CALL plugin (GET /api/plugins/LLM_CALL answers 200; see Plugin overview). Add -H "X-API-Key: $YUKTA_API_KEY" if your server has keys.

Steps

Give the server your key

A key is never written into the workflow. Write it as decrypted:<name>; the default provider reads the environment variable YUKTA_SECRET_<NAME> of the server. Set it where the server starts:

export YUKTA_SECRET_ANTHROPIC=sk-ant-...

Local Ollama needs no key.

Add the LLM_CALL node

provider is openai (set baseUrl for any OpenAI-compatible service such as Groq, vLLM or a gateway; openai-compatible is an alias), anthropic or ollama (baseUrl defaults to the local server). Save as session.json:

session.json

{
  "sessionId": "ask",
  "name": "Ask",
  "workflows": {
    "summarise": {
      "workflowId": "summarise",
      "description": "Summarise a text in two sentences",
      "nodes": [
        { "nodeId": "in", "type": "API_TRIGGER", "config": {} },
        {
          "nodeId": "model",
          "type": "LLM_CALL",
          "config": {
            "provider": "anthropic",
            "model": "claude-sonnet-5-5",
            "apiKey": "decrypted:anthropic",
            "system": "You summarise text in two sentences.",
            "prompt": "${payload.text}",
            "maxTokens": 256,
            "timeout": 60,
            "retries": 1
          }
        },
        { "nodeId": "out", "type": "CONSOLE_TERMINAL", "config": {} }
      ],
      "edges": [
        { "source": "in", "target": "model" },
        { "source": "model", "target": "out" }
      ]
    }
  }
}

The prompt is built from the message with ${payload.text}. timeout is in seconds (default 120), retries retries with backoff (default 0), temperature and maxTokens are optional. For local Ollama use "provider": "ollama", "model": "llama3.2" and drop apiKey.

curl -X POST localhost:9858/api/sessions -H 'Content-Type: application/json' -d @session.json

Start it with a prompt

curl -X POST localhost:9858/api/workflow/start -H 'Content-Type: application/json' \
  -d '{"sessionId":"ask","workflowId":"summarise","payload":{"text":"Yukta runs workflows as flowcharts. ..."}}'
# EXEC=<executionId>

Add a fallback model

A failed call fails the node by default. With failureMode: CONTINUE and routeByStatus: true the node emits on success or failure, so the failure port can lead to a second LLM_CALL with another model:

{ "nodeId": "model", "type": "LLM_CALL",
  "config": { "provider": "anthropic", "model": "claude-sonnet-5-5", "apiKey": "decrypted:anthropic",
              "prompt": "${payload.text}", "failureMode": "CONTINUE", "routeByStatus": true } }

{ "source": "model", "target": "out", "sourcePort": "success" },
{ "source": "model", "target": "fallback", "sourcePort": "failure" }

Verify it worked

curl -s -H 'Accept: application/json' localhost:9858/api/sessions/ask/executions/$EXEC/results

The result from out has the answer in payload.content and token counts in payload.usage, and the run ends SUCCESS. If it fails, GET /api/executions/$EXEC shows the model node at FAILURE; the log names the provider error (a wrong key, an unreachable baseUrl, a timeout).

Variations

Several turns of history. Pass a messages expression instead of a single prompt: Keep session memory.
Tools. Offer the model tools with tools: Use tools and MCP servers.
Spending limits. Add a budget: Set token budgets.
Server-wide guardrails. Per-provider rate limits (yukta.llm.rate-limit.<provider>) and prompt or answer filters (yukta.llm.deny-input-patterns, yukta.llm.deny-output-patterns, yukta.llm.max-input-chars) are server settings: Configuration.

Text from users, web pages or tools goes straight into the prompt. Treat the model’s answer as untrusted input to whatever you wire after it.

AI plugins reference for every LLM_CALL option.
AI and deterministic work for where a model belongs in a workflow.
Use variables and secrets.

Build workflows

Work with AI

Supervise runs

Run and operate

Extend

Steps

Verify it worked

Variations

Steps

Verify it worked

Variations

Related