Configure the LLM provider

Eclipse Atesor™ uses one LLM provider for a run. You select the provider with LLM_PROVIDER and give its API key in the environment or in a .env file. This page lists the providers, tells where the .env file goes, and shows the model of each agent role.

Which providers can I use?

LLM_PROVIDER	Key variable	Models
gemini (default)	GOOGLE_API_KEY	gemini-flash-lite-latest for every role
openai	OPENAI_API_KEY	gpt-4o and gpt-4o-mini
openrouter	OPENROUTER_API_KEY	Free models (ids that end in :free), with a fallback chain
anthropic	ANTHROPIC_API_KEY	claude-haiku-5-5 and claude-sonnet-5-5

Note

The Anthropic provider is in the main branch from 2026-10-15. The v0.2.0 .deb package (2026-10-01) does not include it. To use it, run Atesor from a source checkout until the next release.

The Anthropic provider needs the langchain-anthropic package. pip install -r requirements.txt installs it. If the package is not installed, the key check stops the port with ERROR: langchain-anthropic is not installed (pip install -r requirements.txt).

How do I select the provider?

Set the provider and its key, in your .env file or in the environment:

LLM_PROVIDER=openrouter
OPENROUTER_API_KEY=<your key>

Atesor applies these rules:

• If LLM_PROVIDER is not set, Atesor uses gemini.
• Atesor changes the value to lower case before it uses it.
• An unknown value stops the port with ERROR: Unknown provider: <value>.
• A key that is not set, or that has the placeholder value your_key_here, stops the port with ERROR: <KEY_NAME> not found in environment.

Atesor checks the key only before a port. --setup-only, --probe and a recipe cache hit do not need a key. After a good check, Atesor prints the provider and the models of the supervisor (S), planner (P) and builder (B) roles.

Where does the .env file go?

Install	Put the file here
Source checkout	.env in the root of the checkout
.deb package	~/.config/atesor-ai/.env

For a source checkout, make the file from the template in the checkout root:

cp .env-example .env

For the .deb package, copy the template that the package installs:

mkdir -p ~/.config/atesor-ai
cp /opt/atesor-ai/app/.env-example ~/.config/atesor-ai/.env

Then open the file and set LLM_PROVIDER and the key. The template lists the provider keys and the main optional settings. Environment variables lists every variable.

You can also export the variables in your shell:

export LLM_PROVIDER=gemini
export GOOGLE_API_KEY=<your key>

In which order does Atesor read the files?

A variable that is already in the environment wins over every file. Between the files, the first file that sets a variable wins. Atesor reads them in this order:

1. A .env file in the folder of main.py, or in a parent folder of it.
2. $ATESOR_HOME/.env, when ATESOR_HOME is set.
3. ~/.config/atesor-ai/.env, or $XDG_CONFIG_HOME/atesor-ai/.env when XDG_CONFIG_HOME is set.
4. ~/.local/share/atesor-ai/.env, or $XDG_DATA_HOME/atesor-ai/.env when XDG_DATA_HOME is set.

Caution

With the .deb package, a .env file in your current folder has no effect. main.py calls load_dotenv() without a path, so python-dotenv searches from the folder of main.py (/opt/atesor-ai/app) and its parents. It does not search the current folder. Use ~/.config/atesor-ai/.env, or export the variables.

Which model does each agent use?

The model table is MODEL_CONFIG in src/models.py.

Role	gemini	openai	openrouter	anthropic
supervisor	gemini-flash-lite-latest	gpt-4o-mini	qwen/qwen3-coder:free	claude-haiku-5-5
planner	gemini-flash-lite-latest	gpt-4o	openai/gpt-oss-120b:free	claude-haiku-5-5
scout	gemini-flash-lite-latest	gpt-4o	qwen/qwen3-coder:free	claude-haiku-5-5
builder	gemini-flash-lite-latest	gpt-4o-mini	qwen/qwen3-coder:free	claude-sonnet-5-5
fixer	gemini-flash-lite-latest	gpt-4o	openai/gpt-oss-120b:free	claude-sonnet-5-5
summarizer	gemini-flash-lite-latest	gpt-4o-mini	meta-llama/llama-3.3-70b-instruct:free	claude-haiku-5-5

The gemini, openai and openrouter entries set a temperature. It is 0.0 for the supervisor and the builder, 0.1 for the planner, the scout and the summarizer, and 0.2 for the fixer.

The two Anthropic models do not accept a temperature, so Atesor sends none. In its place, each anthropic entry sets an effort value for the adaptive thinking of the model. The value is low for the supervisor and the summarizer, and medium for the other roles.

These roles make the LLM calls:

• The analyst uses the planner entry.
• The scout uses the scout entry.
• The fixer uses the fixer entry.
• The summarizer writes the recipe, and the artifact curator ranks the build outputs. Both use the summarizer entry.
• In performance mode, the optimizer agent uses the builder entry to write a harness, and the planner entry to propose a program run. If no model of that role answers in time, Atesor asks the models of the other role one time.

The supervisor routes by rules, so its entry makes no calls. The builder entry makes calls only in performance mode.

How do I change a model?

1. Use a source checkout. The .deb package keeps its install tree identical to the package contents.
2. Edit the model value of the role in MODEL_CONFIG in src/models.py.
3. If the new model is a paid model, add its price to MODEL_PRICES_PER_MTOK in the same file.

Without a price entry, Atesor prices a paid model at $3.00 for each million input tokens and $12.00 for each million output tokens. This rate counts toward the cost cap of $1.00 for each port. See Cost control.

How does the OpenRouter fallback chain work?

OpenRouter is the only provider with fallback models. Gemini, OpenAI and Anthropic use the primary model only.

• The chain. OPENROUTER_FALLBACK_MODELS sets the chain as a comma-separated list of model ids. Atesor always adds openrouter/free, the OpenRouter router for free models, at the end. Without the variable, the chain is qwen/qwen3-coder:free, openai/gpt-oss-120b:free, openai/gpt-oss-20b:free, meta-llama/llama-3.3-70b-instruct:free, qwen/qwen3-next-80b-a3b-instruct:free and nvidia/nemotron-3-super-120b-a12b:free.
• Server-side fallback. Each request names three models for OpenRouter to try: the first two models of the chain that are not the primary model, and openrouter/free. OpenRouter refuses a request with more than three.
• Client-side rotation. When a call fails with a provider error, Atesor tries the next model of the chain. For a short Retry-After of 65 seconds or less, it waits one time for each model before it rotates.
• Daily cap. When the account reaches the OpenRouter daily cap for free models, Atesor stops the rotation and uses its deterministic fallback.

The default chain does not contain openrouter/auto, because that router selects paid models only.

Which other LLM settings exist?

Variable	Default	Effect
ATESOR_MAX_OUTPUT_TOKENS	8192	The output token limit of each LLM request
ATESOR_LLM_TIMEOUT	120	The request timeout in seconds. Atesor also stops a call after this value plus 15 seconds (60 seconds for the artifact curator, 240 seconds for a harness of the optimizer agent).
OPENROUTER_FALLBACK_MODELS	The default chain	The OpenRouter fallback chain

Environment variables lists every variable.

Next step

Port your first package.