AI Connectors and Agents

A model call, or an assistant that can act

Your broker already moves sensor data, drives equipment, and stores history. Two LoT pieces add a language model to that world. A TYPE LLM connector sends one prompt and returns text. Actions use it for a classification, a summary, or a short conversation. A DEFINE AGENT is a named assistant that runs on that connector. It can read topics, publish, run connector events, and hand work to other agents — only where you grant it. The connector holds the provider, the model, the key, and the price. Every agent that names it with USES LLM CONNECTOR picks up a change to that connector. An agent is its own LoT entity. It is deployed with -addAgent, and it does not appear in the connector list.

Like a phone line and the people who answer it. The TYPE LLM connector is the line to the model. Each DEFINE AGENT is a person you hire, with a job description and a set of keys. CALL LLM dials the line directly. CALL AGENT asks a specific person to do the job.

When to use which

Prompt in, text out: a TYPE LLM connector and CALL LLM inside an Action. This page.
An assistant that reads or changes the broker: a DEFINE AGENT on that connector. See Agents Overview.

Both need a Growth or Enterprise license, the same row as LLM / Agent on Licensing.

First time? Paste a working agent in Creating Agents in Coreflux. Come back here for providers and CALL LLM. Grants, topics, and the move off TYPE AGENT are in Agent Syntax.

The example below is complete. Deploy the LLM connector with -addConnector.

Quick Start

Ask a model from an Action

One prompt, one answer. The connector has no tools.Store the key once with KEEP SECRET "OPENAI_API_KEY" WITH "sk-...", or with -setSecret. See Secrets.

DEFINE CONNECTOR PlantModel WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "openai"
        WITH MODEL "gpt-6-luna"
        WITH API_KEY GET SECRET "OPENAI_API_KEY"
        WITH API_SURFACE "auto"

DEFINE ACTION ExplainAlarm
ON TOPIC "alarms/raw/+" DO
    CALL LLM "PlantModel.generate"
        WITH (prompt = PAYLOAD)
        RETURN AS {result}
    PUBLISH TOPIC "alarms/explained" WITH (GET JSON "response" IN {result} AS STRING)

Publish an alarm payload to alarms/raw/line3. The explanation arrives on alarms/explained. Read success in {result} before you trust response.

For agent examples, see Agent Examples.

The LLM connector

DEFINE CONNECTOR … WITH TYPE LLM is the connector (DEFINE ROUTE and the -addRoute / -removeRoute commands still work as aliases). Actions call it with CALL LLM. Agents name it with USES LLM CONNECTOR. The connector itself never calls tools.

Picking a provider

Provider	`WITH PROVIDER`	Typical `BASE_URL`	Auth
Ollama (default)	`ollama`	`http://localhost:11434`	none
OpenAI	`openai`	omit (the provider default)	`API_KEY`
Anthropic	`anthropic`	omit (the provider default), or a proxy	`API_KEY`
OpenAI-compatible	`compatible`	your gateway URL	`API_KEY` when that server requires one

Local Ollama. Install Ollama on the broker host, or on a machine the broker can reach, and pull a tools-capable model before you deploy (ollama pull llama3.2 or ollama pull qwen3). The MODEL value must match ollama list. Only some Ollama models call tools. Coreflux connects to the endpoint you set. It does not install Ollama or download models.If the broker runs in Docker, localhost inside the container is the container. Point BASE_URL at the Ollama service on the Compose network (for example http://ollama:11434), or at http://host.docker.internal:11434 when Ollama runs on the Docker host.

Ollama
OpenAI
Anthropic
OpenAI-compatible

The default provider. Omit PROVIDER and the connector still uses Ollama at http://localhost:11434.

DEFINE CONNECTOR LocalLlama WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "ollama"
        WITH BASE_URL "http://localhost:11434"
        WITH MODEL "llama3.2"

API_SURFACE "auto" uses the Responses API, which current reasoning models such as gpt-6-luna require. Leave TEMPERATURE unset for those models.

DEFINE CONNECTOR Gpt6 WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "openai"
        WITH MODEL "gpt-6-luna"
        WITH API_KEY GET SECRET "OPENAI_API_KEY"
        WITH API_SURFACE "auto"
        WITH REASONING_EFFORT "medium"
        WITH PRICE_INPUT "1.25"
        WITH PRICE_OUTPUT "10"

DEFINE CONNECTOR Claude WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "anthropic"
        WITH MODEL "claude-sonnet-5-5"
        WITH API_KEY GET SECRET "ANTHROPIC_API_KEY"

Use compatible for a gateway that speaks the OpenAI HTTP API. API_SURFACE "auto" on this provider uses Chat Completions. Set API_SURFACE "responses" only when that gateway implements the Responses API.

DEFINE CONNECTOR Gateway WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "compatible"
        WITH BASE_URL "https://llm.example.com/v1"
        WITH MODEL "local-large"
        WITH API_KEY GET SECRET "GATEWAY_API_KEY"
        WITH API_SURFACE "chat_completions"

`LLM_CONFIG`

Parameter	Required	Default	Description
`MODEL`	Yes	—	Model name the provider expects (`llama3.2`, `gpt-6-luna`, `claude-sonnet-5-5`)
`PROVIDER`	No	`ollama`	`ollama`, `openai`, `anthropic`, or `compatible`
`BASE_URL`	No	Ollama: `http://localhost:11434`. Others: the provider default	Required in practice for `compatible`
`API_KEY`	Conditional	—	Required for `openai`, `anthropic`, and most `compatible` endpoints. Use `GET SECRET`
`API_SURFACE`	No	`auto`	`openai` and `compatible` only: `auto`, `responses`, or `chat_completions`. `auto` is Responses for `openai` and Chat Completions for `compatible`
`REASONING_EFFORT`	No	see below	`none`, `low`, `medium`, `high`, or `extra_high`, sent on every call. `none` sends nothing
`TEMPERATURE`	No	Ollama: `0.7`. Others: not sent	`0.0`–`2.0`. Leave it unset for reasoning models, which reject a temperature
`MAX_TOKENS`	No	`2048`	Maximum completion tokens for one call. An agent’s `MAX_TOKENS` is a different budget
`PRICE_INPUT`	No	—	Price per million input tokens. Agent tasks then report `cost` on `run_end` when either price is set
`PRICE_OUTPUT`	No	—	Price per million output tokens
`SYSTEM_PROMPT`	No	empty	Default system message for `CALL LLM`. Agents use their own `SYSTEM_PROMPT`

Models that reason before answering spend part of the connector’s MAX_TOKENS on reasoning they don’t show. With 1024 the answer can come back empty. Set MAX_TOKENS to 4096 or more on the TYPE LLM connector for these models.

Without REASONING_EFFORT, only PROVIDER "openai" on the Responses API sends one: Medium. That default is skipped for older families (gpt-4*, gpt-3*, chatgpt-*), which reject it. A compatible connector sends no reasoning effort unless you set it. Set REASONING_EFFORT "none" when an OpenAI model rejects the default. MCP_ROUTES is legacy. On a TYPE LLM connector the broker ignores it and logs a warning to use ADD TOOL. The connector still starts. A TYPE LLM call runs no tools. On DEFINE AGENT it is a legacy clause, and -addAgent refuses it. To give an agent an MCP tool, see Agents and MCP Tools. See Coming from TYPE AGENT. The connector does not probe the provider at start. A wrong key, URL, or model shows up on the first call, as "success": false.

Ollama context and timeouts

Ollama uses the model’s default context window (often 2k–8k tokens) unless the request asks for more, and silently drops the start of a longer prompt. The broker therefore sends num_ctx on every Ollama call, for CALL LLM and for each step of an agent’s tool loop:

It estimates the prompt (messages, tool results, and tool schemas) plus MAX_TOKENS.
It uses 16,384 tokens when that fits, otherwise MAX_CONTEXT_WINDOW. Only two sizes, because each change makes Ollama reload the model.
It never goes above the model’s own maximum, read from Ollama and cached for five minutes.
A prompt that still doesn’t fit fails with context_overflow and is never sent.

Parameter	Default	Description
`CONTEXT_WINDOW`	sized per call	A fixed `num_ctx` in tokens instead of the automatic size. Still capped at the model’s maximum
`MAX_CONTEXT_WINDOW`	`32768`	The largest `num_ctx` the broker picks. Larger windows use more memory on the Ollama host
`FIRST_OUTPUT_TIMEOUT`	`120`	Seconds from sending a request to its first streamed output, model load included. `0` turns it off
`IDLE_TIMEOUT`	`60`	Longest gap, in seconds, between streamed chunks. `0` turns it off. Time spent in tools never counts

These four settings apply to PROVIDER "ollama" only.

DEFINE CONNECTOR LocalLlama WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "ollama"
        WITH BASE_URL "http://ollama:11434"
        WITH MODEL "qwen3"
        WITH MAX_CONTEXT_WINDOW 16384
        WITH FIRST_OUTPUT_TIMEOUT 300

Both timeouts apply per gap, so a long answer that keeps arriving never times out. An agent’s TASK_TIMEOUT and cancel also end a call that is waiting on Ollama. A large model loading cold on a slow machine can take more than two minutes: raise FIRST_OUTPUT_TIMEOUT for it. On a low-memory Ollama host, a 32k window may not fit next to the model. Lower MAX_CONTEXT_WINDOW (for example to 16384).

Failure code	Meaning
`llm_timeout`	No output within `FIRST_OUTPUT_TIMEOUT`, or a gap longer than `IDLE_TIMEOUT`
`llm_unreachable`	Nothing listens at `BASE_URL`: Ollama isn’t running, or the host or port is wrong
`model_not_found`	Ollama doesn’t have `MODEL`. Run `ollama pull` or fix the name
`context_overflow`	The prompt doesn’t fit the largest window the connector allows

Agents report these as the task’s code. CALL LLM returns "success": false with the same message, which names the model and the URL, in error.

Call the model from an Action

Method	What it does
`generate`	One turn. No history
`chat`	Multi-turn. Pass `use_history = "true"` and a `session_id` to keep the thread
`clear_history`	Drops the in-memory thread for that `session_id`

CALL LLM and INVOKE LLM are the same call.

DEFINE CONNECTOR PlantModel WITH TYPE LLM
    ADD LLM_CONFIG
        WITH PROVIDER "openai"
        WITH MODEL "gpt-6-luna"
        WITH API_KEY GET SECRET "OPENAI_API_KEY"

DEFINE ACTION ContinueThread
ON TOPIC "assistants/+/message" DO
    SET "user_id" WITH TOPIC POSITION 2
    CALL LLM "PlantModel.chat"
        WITH (prompt = PAYLOAD, use_history = "true", session_id = {user_id})
        RETURN AS {result}
    PUBLISH TOPIC "assistants/" + {user_id} + "/reply" WITH (GET JSON "response" IN {result} AS STRING)

A success payload for generate and chat includes success, response, model, session_id, used_history, and execution_time_ms. A failure sets success to false and puts the reason in error. History for CALL LLM stays in memory on that connector: the last 50 messages per session_id, dropped when the connector stops or the broker restarts. Agent conversations are separate and survive a restart. See Working memory. You can also fire the connector with TRIGGER "PlantModel" TO {prompt} FOR "generate" (or FOR "chat" for the history mapping). That path publishes the result to the mapping’s DESTINATION_TOPIC, or to {RouteName}/llm/response when that mapping has none. TRIGGER keeps chat history only when the mapping name is chat, unless the JSON payload sets use_history. Status is the usual connector topic $SYS/Coreflux/Routes/{Name}/status. The connector does not publish …/Capabilities or …/Metrics. -discoverAIProviders probes local and cloud model lists and retains a summary on $SYS/Coreflux/AI/Providers. That probe is independent of the providers a TYPE LLM connector accepts.

Agents

DEFINE AGENT has its own section: Agents Overview, Agent Syntax (properties, ADD TOOL, delegation, topics, CALL AGENT, management), and Agent Examples. Moving off TYPE AGENT? See Coming from TYPE AGENT.

Troubleshooting

Ollama never calls a tool

The model does not support tools. Pull a tools-capable tag such as llama3.2 or qwen3.

OpenAI returns HTTP 400

Reasoning models reject TEMPERATURE. Leave it unset. gpt-6-luna rejects a Chat Completions call. Use API_SURFACE "auto" or "responses".

Limitations

CALL LLM waits for the full answer and does not stream. CALL LLM history is in memory only. For agent limits, see Agents Overview.

Next steps

Creating Agents in Coreflux

Paste a site guide, a scheduled report, alert triage, or a shift handover and try them over MQTT.

MCP Connectors

Call external MCP tools from Actions. Agents do not receive those tools automatically.

Quick Start

Coreflux MQTT Broker

LoT Language

Coreflux HUB

VS Code - LoT Notebooks

Coreflux Portal

Using AI with Coreflux

Product Updates

A model call, or an assistant that can act

When to use which

Quick Start

The LLM connector

Picking a provider

`LLM_CONFIG`

Ollama context and timeouts

Call the model from an Action

Agents

Troubleshooting

Limitations

Next steps

Creating Agents in Coreflux

MCP Connectors

​A model call, or an assistant that can act

​When to use which

​Quick Start

​The LLM connector

​Picking a provider

​LLM_CONFIG

​Ollama context and timeouts

​Call the model from an Action

​Agents

​Troubleshooting

​Limitations

​Next steps

Creating Agents in Coreflux

MCP Connectors

A model call, or an assistant that can act

When to use which

Quick Start

The LLM connector

Picking a provider

`LLM_CONFIG`

Ollama context and timeouts

Call the model from an Action

Agents

Troubleshooting

Limitations

Next steps