Provider guide

Cloud API or local model—the bot stays yours.

The provider setting controls the request format. Environment variables supply private credentials and can override the configured base URL without regenerating the project.

Hosted providers

Provider	Environment variable	Notes
OpenAI	`OPENAI_API_KEY`	Uses the documented Chat Completions endpoint for broad compatible-provider support.
Anthropic	`ANTHROPIC_API_KEY`	Uses Anthropic Messages with the system prompt separated.
Google Gemini	`GEMINI_API_KEY`	Uses Gemini generateContent, system instruction fields, and the API-key header.

Model identifiers change over time. Confirm the selected model exists in your provider account before deployment. Provider usage is billed directly by that provider; AMT does not proxy ordinary model conversations. See the official OpenAI Chat Completions, Anthropic Messages, and Gemini generateContent references.

Ollama

ollama pull llama3.2
ollama serve
# generated .env
LLM_BASE_URL=http://127.0.0.1:11434/v1

Ollama’s official OpenAI compatibility endpoint requires an API-key value in some clients but ignores it locally; the runtime supplies a harmless placeholder when no key is set.

LM Studio

Download a model in LM Studio.
Load it and start the local server.
Use the identifier exposed by LM Studio in the Builder’s Model field.
Keep http://127.0.0.1:1234/v1 when bot and model run on the same machine.
For a private remote machine, bind LM Studio to its network interface and use its Tailscale address, such as http://100.x.y.z:1234/v1. Do not expose an unauthenticated model server to the public internet.

OpenWebUI and compatible servers

Select OpenAI-compatible endpoint. Set the exact model name, enter the server’s /v1 base URL, and place its key in LLM_API_KEY. For OpenWebUI pipelines, the model and authentication requirements are defined by that OpenWebUI installation.

LLM_BASE_URL=https://models.example.com/v1
LLM_API_KEY=replace_with_private_key

Remote endpoint warning.

A non-local endpoint receives the bot’s system prompt, recent memory, user message, and relevant knowledge context. Use only providers you trust and review their retention terms.

Tuning

Setting	Effect
Temperature	Higher values increase variation; lower values are steadier.
Max output tokens	Caps each response. Reasoning models may consume hundreds of hidden reasoning tokens before visible text; use at least 500–1,200 when a small cap returns no visible answer.
Context size	Documents intent for the selected model; actual limits remain provider-controlled.
Remembered messages	Controls how many prior local messages are added to each request.

← Platforms Deployment →