AI models

llm and agent steps — and the workflow assistant — run on AI models. Every account chooses its own mix:

	Murmurator AI	Your own keys
Setup	One click	Add a provider API key
Models	Kimi K3, DeepSeek V4 Pro, GLM-5.3, DeepSeek V4.1 Flash	Anything your provider offers
Billing	Per token on your Murmurator invoice	Directly by your provider, nothing added
Limits	Monthly spend limit you control	Your provider's limits
Data	Processed by our inference provider (DigitalOcean)	Under your own provider agreement

Named models

Every model in an account has a short name — smart, fast, reviewer — and steps refer to that name:

- key: review
  kind: llm
  model: smart
  prompt: "Review this diff: {{ steps.pr.output.diff }}"

• One model is the account default, used by any step without a model.
• Names decouple workflows from vendors. Point smart at a different model and every workflow using it switches over, with no edits.
• Built-in and own-key models live side by side, so one workflow can use Murmurator AI for one step and your own OpenAI key for another.
• Saving a workflow fails if a step names a model the account doesn't have.

Only owners and admins can add, rename or remove models.

Murmurator AI

Built-in open-weight models, no provider account needed. Add them under AI models → Murmurator AI with Use; each becomes a named model (named after its key, such as balanced, which you can rename).

Model	Best for	Input	Output
Kimi K3 (max)	The hardest reasoning and long agent work	$3.90	$19.50
DeepSeek V4 Pro (deep)	Deep analysis, careful reviews, multi-step agents	$2.26	$4.52
GLM-5.3 (balanced)	The balanced default for most steps	$1.82	$5.72
DeepSeek V4.1 Flash (fast)	Classification, extraction and summaries	$0.39	$1.56

Prices are per 1 million tokens; cached input is priced at a reduced rate.

Every model response in a step is recorded with its input and output tokens and what it would cost at those rates, linked to the run and step. See it under AI models → Usage & limits.

Your own providers

Add a provider once under AI models → Your providers with its API key, then add models from it with the provider's model ID:

Provider	Example model IDs
Anthropic	claude-sonnet-5, claude-haiku-4-5
OpenAI	gpt-5, gpt-5-mini
Google Gemini	gemini-3-pro
OpenRouter	anthropic/claude-sonnet-5 and hundreds more
Mistral, DeepSeek, xAI	The provider's model IDs
Ollama	Any model on your server (set the API base URL)

Keys are encrypted at rest and never shown again after saving. An optional API base URL supports gateways and proxies. Calls go straight to your provider under your terms, and Murmurator never adds a markup.

For Ollama, set the API base URL to your server, such as https://ollama.example.com (Murmurator adds /v1 for you). Ollama's default context window of 4,096 tokens is too small for agent steps and the assistant, so raise it with OLLAMA_CONTEXT_LENGTH (16,384 or more) or a model's num_ctx parameter. When a model runs out of room, the step fails with an error saying so.

A model you host yourself can do everything a hosted one can: llm and agent steps, the assistant, and the coding agent in a workspace. The coding agent asks the most of it — thousands of tokens of instructions every turn — so give it your largest context window and expect it to take minutes rather than seconds.

How models are used in runs

• llm steps send one request. With a schema, the model must reply with JSON matching it, available as output.data.
• agent steps send a request per turn and call tools between turns, so they use more tokens — max_iterations caps how many turns they take.
• Each step's log names the model it used, and the run records token totals.
• A step fails with a clear error if its model isn't configured, a provider rejects the request, or a Murmurator AI limit is reached. See Runs.

Choosing a model:

• Start with GLM-5.3 (or your provider's equivalent) for reviews, summaries and agents.
• Use DeepSeek V4.1 Flash for high-volume classification, extraction and short summaries.
• Save DeepSeek V4 Pro and Kimi K3 for steps where quality clearly matters more than cost, such as complex analysis or long agent loops.

The workflow assistant

The assistant that builds your workflows uses your account's default model. If the account has no models yet, it runs on Murmurator AI's built-in model.

Either way, the assistant is included in your plan and is never billed as Murmurator AI usage, even when the default is a Murmurator AI model. A generous monthly fair-use allowance applies. If your default model uses your own key, your provider bills you for those tokens as usual.