These examples use the catalog, keeping model names and routes explicit. Run
catalog-refresh first when provider inventories may have changed.
uv run python -m most catalog-refresh --show-models
uv run python -m most catalog-options --capability chat
uv run python -m most catalog-options --capability embedding
uv run python -m most catalog-options --capability image
uv run python -m most catalog-options --capability speech
uv run python -m most catalog-options --capability transcriptionTypical input/output mappings are:
text -> text ordinary chat or coding
text,image -> text image understanding
text -> embedding search/RAG vector
text -> image image generation
text -> audio speech synthesis
audio -> text transcription, for example Whisper
uv run python -m most ai-chat \
--provider ollama --model granite4.1:3b \
"Summarize the idea of a database in one sentence."uv run python -m most ai-chat \
--provider anthropic --model claude-sonnet-5 --route api \
"Review this function for a possible bug."
uv run python -m most ai-chat \
--provider google --model gemini-3.5-flash --route api \
"Give me three names for this project."Use --route auto (the default) to let MOST select an available route.
For a local vLLM, LM Studio, or other OpenAI-compatible service, add its
endpoint as an openai-compatible catalog route. Localhost/private endpoints
can be audited without an API key; cloud endpoints still require their
configured credential:
uv run python -m most catalog-audit --provider ollama --show-modelsuv run python -m most cli-chat claude --allow-unknown-connectivity "Explain this error."
uv run python -m most cli-chat agy --allow-unknown-connectivity "Suggest a simpler implementation."These use authentication configured in the installed CLI; MOST does not need the CLI subscription token.
uv run python -m most ai-embed \
--provider google --model models/gemini-embedding-001 \
--input examples/sample.txt --output /tmp/sample.embedding.jsonThe output is JSON containing the model, vector dimensions, and values. Embeddings are for semantic search and RAG, not chat responses.
uv run python -m most ai-image \
--provider google --model models/gemini-3-pro-image-preview \
--output /tmp/most-example-image.bin \
"A clean blue geometric logo for a command-line AI tool"The command prints the returned MIME type. Image-generation models are not
valid ai-chat models.
uv run python -m most ai-speech \
--provider google --model models/gemini-2.5-pro-preview-tts \
--output /tmp/most-example-speech.bin \
"Welcome to the MOST practical example."This is text-to-speech. Whisper has the opposite direction: audio input and text output.
For OpenAI Whisper transcription:
uv run python -m most ai-transcribe \
--provider openai --model whisper-1 \
--input examples/media/sample-speech.wavThe transcription is printed and recorded in the session journal.
Refresh provider inventories and recheck models after a failed unified chat:
uv run python -m most catalog-refresh --show-models
uv run python -m most catalog-healthThe same transcription and embedding interfaces can use an OpenAI-compatible local or institutional endpoint when the selected catalog model advertises the required modality, for example:
uv run python -m most ai-embed \
--provider ollama --model embeddinggemma:latest \
--input examples/sample.txt --output /tmp/local.embedding.jsonTask records include provider usage when returned and estimate cost from the catalog pricing metadata when both are available.
uv run python -m most ai-image-analyze \
--provider google --model gemini-3.5-flash \
--input examples/media/sample-public-domain.jpg \
"Describe the important objects and colors."MOST rejects this request if the selected model does not advertise image
input. The same protection prevents sending an embedding model to ai-chat.
Every successful chat prints a session ID. Inspect an execution with:
uv run python -m most inspect-execution EXECUTION_ID
uv run python -m most list-sessionsBrowser profiles are available for Gemini, ChatGPT, Claude, and CERIT. Create one named profile per account; each profile keeps its own login cookies:
uv run python -m most browser-chat gemini --profile gemini-edu
uv run python -m most browser-chat gemini --profile gemini-personal
uv run python -m most browser-chat claude --profile claude-work
uv run python -m most browser-chat chatgpt --profile chatgpt-personal
uv run python -m most browser-chat cerit --profile cerit-institutionLog into each profile once when Firefox opens. API and CLI routes use their own credential systems and do not need browser profiles.
uv run python -m most credentials set openai
uv run python -m most credentials set anthropic
uv run python -m most credentials set google
uv run python -m most credentials set einfraIf a standard environment variable is already set, copy it without displaying the secret:
uv run python -m most credentials set google --from-envKeys are stored in the operating system keyring, not in the catalog.