Skip to content

Providers#

--model (or BLYTE_MODEL) names the model, and the name picks the provider. blyte reads the credentials itself; they are removed from the environment of every command the model runs.

model name provider credentials
gemini-3.8-flash-high (default) Google Gemini; -low / -medium / -high sets the thinking level GEMINI_API_KEY (or GOOGLE_API_KEY)
vertex/<model> (e.g. vertex/gemini-4-argon) Google Gemini on Vertex AI; ADC authentication gcloud auth application-default login, GOOGLE_CLOUD_PROJECT
claude-… or anthropic/… Anthropic ANTHROPIC_API_KEY
openai/<model> any OpenAI-compatible API: OpenAI, or a local vLLM or LM Studio server OPENAI_API_KEY, OPENAI_BASE_URL (default https://api.openai.com/v1)
openrouter/<vendor>/<model> OpenRouter OPENROUTER_API_KEY
<name>:<tag> or ollama/… Ollama, local or cloud; thinking off unless the name ends in +think OLLAMA_HOST (e.g. http://localhost:11434) or OLLAMA_API_KEY
chatgpt/<model> (e.g. chatgpt/gpt-5.5) a ChatGPT subscription, through the Codex CLI's login ~/.codex/auth.json from codex login ($CODEX_HOME/auth.json when CODEX_HOME is set)

Choosing an implementer model#

In the plan → delegate → review pattern, blyte's model does the implementation, so it should be cheap, fast and good at following precise instructions. It does not need to be good at planning: the frontier model above it does that.

model status
Gemini 3.8 Flash (default) measured: every paired run and the FrontierHarness run used it. Start here.
Claude Haiku, GPT mini-class models supported (via claude-…, openai/…, openrouter/…), not yet measured as implementers
local models via Ollama supported; qwen3:14b handled small tasks but failed a real repository task in our tests. Worth trying on simple, well-tested tasks; slow on a laptop.

Gemini 3.8 Flash and Gemini 4 Argon (also through vertex/) have list prices in blyte today, so they are the only models for which the report's cost_usd and --max-cost work. For another model, compare token counts in the report.

Notes per provider#

Gemini. The default. The thinking suffix is part of the name: gemini-3.8-flash-low thinks least and costs least. Models prefixed with vertex/ (e.g. vertex/gemini-4-argon) reach Gemini through Vertex AI using Google Cloud Application Default Credentials (gcloud auth application-default login) rather than an API key. The project is read from GOOGLE_CLOUD_PROJECT (or CLOUDSDK_CORE_PROJECT), and location from GOOGLE_CLOUD_LOCATION (default global). Tokens are cached for 45 minutes and refreshed automatically. Gemini calls stream over SSE (:streamGenerateContent?alt=sse) with thought summaries on (thinkingConfig.includeThoughts: true) so chunks arrive while the model thinks; summaries are discarded and not shown. Set BLYTE_GEMINI_THOUGHT_SUMMARIES=0 to turn them off.

Anthropic. Prompt caching is on: blyte marks cache breakpoints on the tool definitions and the latest message, so each turn pays full price only for what is new.

OpenAI-compatible. Anything that speaks /v1/chat/completions. Point OPENAI_BASE_URL at a local server to run open models without a key.

Ollama. A local model needs its context window set above blyte's fold budget. Start the server with a larger context and lower the budget:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve
blyte --model qwen3:14b --fold-budget 16000 "..." --gate "..."

Thinking is off by default (it costs turns of time on small models); add +think to the name to turn it on.

ChatGPT subscription. Uses the login the Codex CLI stores in ~/.codex/auth.json. blyte only reads it: it never refreshes or rewrites the token. If the token is about to expire, blyte stops with an authentication error and asks you to run any codex command to refresh it. Usage counts against your ChatGPT plan's limits; when a plan's quota is exhausted, blyte stops at once and says when it resets.

Retries and quotas#

Burst rate limits (HTTP 429 with a short retry delay), server errors and dropped connections are retried with backoff (2, 4, 8, 16, 32, then 60 seconds, six times), never past --max-seconds. Streamed Gemini calls fail after BLYTE_MODEL_IDLE_TIMEOUT_SECS (default 300 s; Gemini can think silently for over two minutes before a long reply) of silence with no bytes received, or once BLYTE_MODEL_TIMEOUT_SECS (default 1800 s for streamed calls) has passed. Other providers use BLYTE_MODEL_TIMEOUT_SECS as a whole-call HTTP timeout (default 180 s, with a read timeout 30 s shorter, at least 30 s) for slow providers such as a local Ollama instance. An exhausted quota (a daily limit, insufficient_quota, or a reset more than five minutes away) is not retried: the run ends with a message saying when it resets.

Cost#

When blyte knows a model's list price, the report carries cost_usd and --max-cost can cap it. Costs are computed from the provider's reported usage, with cache hits at their discounted price. Today the price table covers Gemini 3.8 Flash and Gemini 4 Argon; for other models cost_usd is null and --max-cost is refused rather than silently ignored. Token counts are reported for every provider.