Models

These settings choose the AI model that the agent uses. tiny-harness works with:

• OpenAI, through the [openai] section.
• Any server that copies the OpenAI API: Ollama, OpenRouter, vLLM, LM Studio and others. Use the [openai] section with base_url. The agent always uses the [openai] section. tiny-harness also has an Anthropic adapter, but the CLI and running_harness do not use it yet. See [anthropic].

Examples

OpenAI:

[openai]
model = "gpt-6.1-sol"
reasoning_effort = "medium"

export OPENAI_API_KEY=sk-...

A local Ollama, with no key:

[openai]
base_url = "http://127.0.0.1:11434/v1"
api = "chat_completions"
model = "qwen3:8b"
context_window_tokens = 16384
timeout = "PT600S"

OpenRouter:

[openai]
base_url = "https://openrouter.ai/api/v1"
api = "chat_completions"
model = "provider/model-name"      # a model name from openrouter.ai/models

export OPENAI_API_KEY=sk-or-...   # your OpenRouter key

[openai] settings

Setting	Default
model	"gpt-6.1-sol"
models	model only
reasoning_effort	none
base_url	"https://api.openai.com/v1"
api	"responses"
context_window_tokens	known for OpenAI models, else 400 000
max_output_tokens	2000
timeout	"PT60S" (60 seconds)

model

The model the agent uses for a new task.

• Default: "gpt-6.1-sol"
• Use a name that your server knows. For Ollama, that is the name you pulled, for example "qwen3:8b".

models

The models that a user can choose with the /model command.

• Default: only model.
• Set it when: you want users to switch models in a task.
• Example: models = ["gpt-6.1-sol", "my-other-model"]
• The list must contain model. An empty list, or a list without model, stops tiny-harness.

reasoning_effort

How much the model thinks before it answers: "low", "medium" or "high". A user can change it in a task with /effort.

• Default: none. tiny-harness sends no effort, and the model uses its own default.
• A model that does not accept an effort can refuse the request. Leave it unset for such a model.

base_url

The address of an OpenAI-compatible server.

• Default: "https://api.openai.com/v1"
• Set it when: you use a server other than OpenAI. With base_url, OPENAI_API_KEY is optional.
• tiny-harness sends requests to this address only. It ignores the OPENAI_BASE_URL environment variable.
• These values stop tiny-harness:
  • An address with a user name or password in it (https://user:pass@...). Put the key in OPENAI_API_KEY.
  • An http:// address to another computer while OPENAI_API_KEY is set. Use https://, so the key is not sent in clear text. http:// is allowed for localhost and 127.0.0.1.

api

Which OpenAI API the server speaks.

• Default: "responses" (the OpenAI Responses API)
• Set it when: your server has only the older Chat Completions API. Then use "chat_completions". Many local servers need this.
• For OpenAI's own reasoning models, keep "responses".

context_window_tokens

The maximum input size of the model, in tokens. When a conversation becomes too long, tiny-harness writes a summary of the older part. It uses this number to decide when.

• Default: tiny-harness knows the size of OpenAI models. For other models it uses 400 000.
• Set it when: you use a model that tiny-harness does not know, or a server that gives the model a smaller window. For Ollama, use the value of OLLAMA_CONTEXT_LENGTH.
• A whole number greater than 0.

max_output_tokens

The maximum length of one model reply, in tokens.

• Default: 2000

timeout

How long tiny-harness waits for one model reply.

• Default: "PT60S" (60 seconds)
• Set it when: your model is slow. A local model can need "PT600S".

When a model call fails because of a timeout, a rate limit or a server error, tiny-harness tries again. See [retries].

[anthropic] settings

tiny-harness reads and checks this section. The CLI and running_harness do not use it yet, so the agent uses [openai]. The Anthropic adapter is not tested end to end.

[anthropic]
model = "claude-opus-5-5"

export ANTHROPIC_API_KEY=...

Setting	Default	What it does
model	"claude-opus-5-5" when only the key is set	The model
models	model only	The models /model can choose. Same rules as [openai] models
reasoning_effort	none	Same as [openai] reasoning_effort