Mankomail

OpenAI-compatible API ​

Link this node to the “Model” port of an AI node to call a server speaking the OpenAI protocol without being one of the named providers: vLLM, LM Studio, an internal proxy. The server address is set in the connections.

OpenAI-compatible API is a model provider, not a step. Place it on the canvas and draw a link from it to the model port of an AI node (Categorize, Extract, Summarize, Free instruction (AI) or Compose (AI)): that node then runs on a server that speaks the OpenAI protocol without being one of the named providers (vLLM, LM Studio, an internal proxy). The provider node is never executed, has no data input or output, and writes nothing to the run data.

An AI node with nothing linked to its model port uses the instance defaults: the model the administrator set for its use (Classify, Extract, Compose, General use) on the Connections page (Artificial intelligence section). Linking a provider is how you choose the AI of one node without changing the others. No API key or secret appears in the workflow: credentials stay in the organisation's OpenAI-compatible API connection.

The server address and the API key are set on the connection, on the Connections page (Artificial intelligence section), not on the node. Use this provider only when none of the named providers fits.

At a glance ​

  • Type: llm.openai_compatible · version 1
  • Category: AI
  • Kind: Model provider — plugged into an AI node, never run as a step
  • Connection: OpenAI-compatible API
  • Provides: model (llm.model)

Connection ​

This node uses a OpenAI-compatible API connection. Without one chosen, it uses the default connection of that provider.

Parameters ​

connection ​

Connection — Which key to use at this provider, when the organisation has several. Left empty, the provider default connection applies.

  • Type: Connection (credential)
  • Required: No

model ​

Model — The model to use. Left empty, the instance default for the connected node's purpose applies. This provider has no fixed catalogue: copy the exact id of a model enabled under Connections › AI.

  • Type: Text (string)
  • Required: No
  • Default: "" (empty)
  • 200 characters at most
  • Expressions: {{ }} not accepted

temperature ​

Temperature — Optional. Left empty, the provider default applies. Some recent models reject this parameter: it is then dropped, never fatal.

  • Type: Number (number)
  • Required: No
  • From 0 to 2

Data produced ​

A model provider is never run: it adds nothing to the run data. The AI node it is plugged into publishes the result.

Example ​

Route an AI step through the company's internal model gateway. Add an OpenAI-compatible API node with:

connection: (empty: the provider's default connection)
model: (a model served by the server and enabled on the connection)
temperature: (empty)

Link it to the model port of the Free instruction (AI) node. Every run of that node now calls this provider and model; the other AI nodes of the workflow are unaffected.

Connection, model and temperature ​

  • Connection: left empty, the provider's default connection applies, so replacing a key is done once, by the administrator. Pick a connection only when the organisation has several connections for this provider (a key re-billed to one client, for instance). Only connections of this provider are offered.
  • Model: left empty, the instance default for the linked node's use applies, at this provider. The editor has no fixed list for this provider: it suggests the models the administrator enabled on the chosen connection; otherwise copy the exact identifier of a model enabled on the Connections page (Artificial intelligence section). The field is not templatable: a {{ }} expression would be sent to the provider as is. The editor flags an identifier that the connection does not allow.
  • Temperature: optional, between 0 and 2. It applies only when the AI node does not set its own. Models that reject this parameter simply ignore it.

Tips ​

  • One provider per port. A second provider linked to the same model port is refused at validation. A provider node linked to nothing raises a warning, not an error.
  • Errors at run time. The AI node, not the provider, fails: llm.provider_not_configured when the instance has no connection for this provider, llm.connection_not_found when the chosen connection was deleted (there is no silent fallback to the default connection), llm.connection_provider_mismatch when it belongs to another provider, llm.model_not_allowed when the model is not enabled, llm.no_default_model when the model is empty and no default is known for that use at this provider. See Error handling.
  • Test runs. In test runs the AI node really calls this provider and model, so the test measures what the published workflow will do. If the instance has no connection for this provider, the output is fabricated and the run detail says the AI was skipped.