AI settings & models
The platform model runs the chat and every unbound AI step. How to bind Anthropic, OpenAI, Gemini or OpenRouter to an agent step, and how spend is metered.
The chat that builds always runs on the MadMax platform model. The "which model?" question only exists for the AI steps inside your workflows.
The MadMax default
Out of the box, everything runs on the MadMax platform model: picked and tuned by us, billed through your workspace's AI credits, no keys required. When an agent step's Provider field reads MadMax (platform default), that's what executes. Today the platform default is luna.
The platform actually keeps a small roster of models by job, and MadMax operators can swap any of them without a deploy:
- Canvas model: every non-BYO canvas turn (the chat, build-mode runs) and every unbound custom, classify and summarize step.
- Vision model: the extract task reading PDFs and images.
- Utility model: background jobs such as contextual chunk headers and reranking in knowledge bases, query rewriting, memory summaries and SMS reply suggestions.
- Docs assistant model: the public ask the docs panel.
- Embedding model: the vectors behind knowledge-base search.
You don't configure any of this. It is listed so you know which model a given step runs on when nothing is bound.
Bringing your own model (BYO)
Connect an AI provider under Settings → connectors. The AI category holds Claude (Anthropic), OpenAI, OpenRouter, Google Gemini and Kimi; each takes an API key. Then, on any AI Agent step:
- Provider: pick the connection. Empty keeps the platform default.
- Model: appears once a Provider is bound. Most providers show
a short menu of ids; OpenRouter shows a free-text field, so
paste the id exactly as openrouter.ai lists it
(
openai/gpt-5.6-luna), routing suffix included (:nitro). Blank on OpenRouter falls back to the platform's canvas model.
Three rules worth knowing:
- The Model field is honored only with a Provider bound. Unbound steps always run the platform default, whatever the field says.
- BYO calls bill to your provider account. The metering layer records the key source, so they never draw down MadMax credits.
- The chat, build verification and embeddings never use your key. BYO is per step, on purpose.
Spend
Platform AI usage meters per call into your workspace ledger, attributed to the workflow and step that made it. Two guards sit on top:
- Credit gate: when the workspace's AI credit is spent, the chat replies with a plan-limit message instead of starting a turn. Sessions on a BYO key are exempt.
- Build checkpoint: the auto-verify walk tracks the AI calls a single build has made. Past its budget it pauses and posts a "Checkpoint" message with the call count (and the cost when the model is priced). Reply with anything and it continues with a fresh budget; the plan and every verified step are already saved.
Settings → billing shows the credits remaining and the pools they come from (plan credits, welcome credits, top-ups). One credit is one cent, and it powers Max and the AI steps alike.
FAQ
Which is better? The default is tuned for building (fast, structured-output-safe, and it is what every verification walk is measured against). BYO makes sense when you have credits elsewhere, a compliance need, or a model preference for a specific agent step.
Do web-search-enabled agent steps work on BYO? No. The live web-search tool is a platform-model feature; on a bound connection the toggle is ignored (the field's hint says so).
Can I set temperature? Yes, per step, under Advanced. Classify pins it to 0 unless you override; the others leave the provider default.