Model

Overview
What it is	A model on a provider, and how agt calls it
CLI	agt model
Agent tools	resource_* with kind: "model"
Web console	Resources → Model
File	models/<category>/<name>/model.yaml
Crates	agt-model, agt-inference

What it does

A model resource gives a name to one model on a provider, with the protocol and settings to use. Native agents refer to a model; changing the model resource changes every agent that uses it from its next run. You can define several model resources for the same underlying model, for example a fast and a careful variant.

agt streams responses, retries transient failures up to 10 attempts, and summarizes old context automatically when a conversation nears the context window.

How to use it

Web console: Resources → New resource → Model. Choose the provider, click Fetch model list (获取模型列表) or type the model ID, and choose the API format.

CLI:

agt model create fast --provider openai --model gpt-5-mini --api-format responses --tool-call true
agt model update fast --reasoning-effort low
agt model update fast --clear-reasoning-effort

Agent tools:

{"kind":"model","reference":"fast","create":true,
 "data":{"provider":"openai","model":"gpt-5-mini","params":{"api_format":"responses","tool_call":true}}}

Common parameters

Parameter	Meaning
--model	The model ID at the provider
--api-format	chat-completions, responses, anthropic-messages or gemini; must match what the provider supports
--tool-call true	Required for agents that use any tool
--reasoning-effort	Thinking level for reasoning models
--context-window	Input tokens the model accepts; set it from the provider's documentation

All parameters

agt model create <name>

Option	Required	Default	Meaning
--provider <ref>	yes		The provider
--model <id>	yes		Model ID at the provider
--api-format <f>	no	chat-completions	chat-completions, responses, anthropic-messages or gemini. agt never switches automatically.
--temperature <t>	no	0.7	0 to 2. Not sent when a reasoning effort is set.
--max-tokens <n>	no	not sent	Output token limit
--timeout <s>	no	60	Seconds to wait for the response to start, for each read, and between output chunks once the reply is streaming; a stalled reply is retried
--reasoning-effort <level>	no	not sent	none, minimal, low, medium, high, xhigh, max, or any value the provider accepts
--tool-call <bool>	no		true if the model supports function calling
--context-window <n>	no	272000	Input token window, 4096 to 2000000. Used for automatic compaction; does not limit output.
--category <c>	no	default	Category

agt model update <name>

Accepts the same options as create except --provider and --model, plus:

Option	Meaning
--clear-max-tokens	Stop sending an output limit
--clear-reasoning-effort	Stop sending a reasoning effort

list, info, delete and category are described in Resources.

File: model.yaml

name: fast
provider: { category: default, name: openai }
model: gpt-5-mini
api_format: responses
params:
  temperature: 0.7
  max_tokens: null
  timeout: 60
  reasoning_effort: null
capabilities:
  context_window: 272000
  tool_call: true

Tool fields (resource_write, kind: "model")

On create: provider (required), model (required), and optional params with any of the fields below. On update, give the fields directly in data.

Field	Meaning
temperature	0 to 2
max_tokens	Positive integer
timeout	Seconds
api_format	chat-completions, responses, anthropic-messages or gemini
reasoning_effort	As for --reasoning-effort
tool_call	Boolean
context_window	4096 to 2000000

clear: ["max_tokens"] or clear: ["reasoning_effort"] removes those settings.

Request behavior

Behavior	Details
Streaming	Requests use streaming. Providers that ignore it still work, without live preview.
Retries	Connection errors, HTTP errors, invalid JSON and empty replies are retried up to 10 attempts in total, waiting from 250 ms up to 4 s. Configuration errors and explicit refusals are not retried.
Compaction	At 90% of the context window, older history is summarized with the same model.
Input	Text only; images are not supported.
Anthropic	Sends x-api-key and anthropic-version: 2023-06-01. Marks the system prompt, tools and newest message for prompt caching. Without max_tokens, asks for up to 16384 output tokens. reasoning_effort becomes an extended-thinking budget (low 2048, medium 8192, high 16384, xhigh 32768, max 64000, or a number); thinking is turned off for the one call where a tool is required.
Gemini	Sends x-goog-api-key. Tool schemas go as JSON Schema (parametersJsonSchema). reasoning_effort becomes a thinking budget (none 0, low 2048, medium 8192, high 16384, max lets the model decide, or a number).
Thinking	Provider reasoning is never shown or stored in records. Thinking blocks and signatures that the provider needs back are kept with the run so the next step can send them; a run suspended on a task saves them in its local checkpoint.

Related modules

• Provider: where requests go
• Agent: agents that use the model
• Permissions: a model can also act as the approval reviewer