Skip to content
Docs

Documentation navigation

AI Providers

OpenAI, Anthropic, Gemini, Bedrock, Mistral, Cohere, and other AI-provider toolsets.

#AI providers

LLMs, multi-modal models, transcription, image generation, and search APIs commonly used by agents.

#Agent-ready behavior (overview)

Across this group, only the list / discovery tools were touched for agent ergonomics; chat / completion / embedding / image / TTS tools are unchanged (they already return content, not lists). For each listing tool you get:

  • A compact summary array with a response-local display ref (model_ref, file_ref, batch_ref, job_ref, store_ref, voice_ref, prediction_ref, deployment_ref, transcript_ref, project_ref, dataset_ref, kb_ref) safe to show in final answers.
  • Raw provider IDs are hidden by default. Pass include_ids=True only when a follow-up tool needs them.
  • Default page size is 20-25 for resource lists (10 for AssemblyAI transcripts and Bedrock knowledge bases). Search APIs (Tavily, Brave, SerpAPI) cap max_results / count / num at 10.

Write tools that take an identifier (get_file, delete_file, create_batch, get_batch, cancel_batch, delete_response, get_vector_store, delete_vector_store, cancel_fine_tuning_job, create_fine_tuning_job, get_prediction, cancel_prediction, get_transcript, delete_transcript, lemur_task) accept the dict returned by the corresponding list / get tool as well as a raw ID string.

For chat-style models, list_models always keeps the model name as model_name because callers need it to invoke the model -- only internal handles (file/batch/job/store/etc.) are hidden.

#OpenAIToolSet

Python 
from maivn_tools import OpenAIToolSet

connector = OpenAIToolSet(api_key=secrets["OPENAI_API_KEY"])

Tools: chat_completion, responses_create, get_response, delete_response (destructive), create_embedding, list_models, get_model, delete_fine_tuned_model (destructive), generate_image, speech, list_files, get_file, delete_file (destructive), create_batch, list_batches, get_batch, cancel_batch (destructive), moderate, create_fine_tuning_job, list_fine_tuning_jobs, cancel_fine_tuning_job (destructive), create_vector_store, list_vector_stores, get_vector_store, delete_vector_store (destructive).

Agent-ready behavior

  • list_models (default limit 20): summary with model_ref, model_name, owned_by, created. Set include_ids=True for raw id.
  • list_files (default limit 20): summary with file_ref, filename, purpose, bytes, created_at, status. Raw file-... IDs hidden; opt in with include_ids=True (needed for delete_file, create_batch, create_fine_tuning_job).
  • list_batches (default limit 20): summary with batch_ref, endpoint, status, request_counts, created_at. Opt in for raw batch_id.
  • list_fine_tuning_jobs (default limit 20): summary with job_ref, model, fine_tuned_model, status, created_at. Opt in for raw job_id.
  • list_vector_stores (default limit 20): summary with store_ref, name, file_counts, status, created_at. Opt in for raw store_id.
  • Tolerant inputs: get_file, delete_file, create_batch, get_batch, cancel_batch, get_vector_store, delete_vector_store, cancel_fine_tuning_job, and create_fine_tuning_job accept the raw ID, the file/batch/job dict returned by the matching list tool, or the full {"data": [...]} list payload.
  • Destructive: delete_response, delete_fine_tuned_model, delete_file, cancel_batch, cancel_fine_tuning_job, delete_vector_store. Confirm with the user before invoking.
Python 
files = connector.list_files()                 # compact summaries
files = connector.list_files(include_ids=True) # also include file_id
connector.delete_file(files["files"][0])       # accepts the dict

#AnthropicToolSet

Python 
from maivn_tools import AnthropicToolSet

connector = AnthropicToolSet(api_key=secrets["ANTHROPIC_API_KEY"])

Tools: create_message, count_tokens, list_models, get_model, create_message_batch, get_batch, cancel_batch (destructive).

Agent-ready behavior

  • list_models (default limit 20): summary with model_ref, model_name, display_name, created_at. model_name is kept because it is also the model identifier callers pass to create_message. Set include_ids=True to also include the raw id.
  • get_batch and cancel_batch accept a raw msgbatch_... string or a batch dict returned by create_message_batch / get_batch.
  • Destructive: cancel_batch. Confirm before invoking.
Python 
models = connector.list_models()
batch = connector.create_message_batch(requests=[...])
connector.get_batch(batch)  # dict accepted

#AzureOpenAIToolSet

Python 
from maivn_tools import AzureOpenAIToolSet

connector = AzureOpenAIToolSet(
    endpoint="https://my-rsrc.openai.azure.com",
    api_key=secrets["AZURE_OPENAI_KEY"],
    api_version="2024-08-01-preview",
)

Tools: chat_completion(deployment, ...), embeddings(deployment, ...), generate_image(deployment, ...), list_models.

Agent-ready behavior

  • list_models (default limit 25): summary with model_ref, model_name, capabilities, lifecycle_status. model_name is kept because it is the deployment identifier callers pass to chat_completion / embeddings / generate_image. Set include_ids=True for the raw id.

#GeminiToolSet

Google AI Studio / Generative Language API.

Python 
from maivn_tools import GeminiToolSet

connector = GeminiToolSet(api_key=secrets["GEMINI_API_KEY"])

Tools: generate_content, count_tokens, embed_content, list_models, get_model.

Agent-ready behavior

  • list_models (default limit 25): summary with model_ref, model_name, display_name, supported_methods, input_token_limit, output_token_limit. model_name is the "models/..." resource name callers pass to generate_content. Set include_ids=True for the raw name.

#BedrockToolSet

AWS Bedrock Runtime + control plane. Pass a SigV4 AuthStrategy.

Python 
from maivn_tools import BedrockToolSet

connector = BedrockToolSet(region="us-east-1", auth=sigv4_auth)

Tools: list_foundation_models, converse, invoke_model, list_knowledge_bases, retrieve.

Agent-ready behavior

  • list_foundation_models (default limit 25): summary with model_ref, model_name, provider, input_modalities, output_modalities, model_lifecycle. model_name is the modelId callers pass to converse / invoke_model. Set include_ids=True for the raw model_arn.
  • list_knowledge_bases (default limit 10): summary with kb_ref, knowledge_base_id (kept because callers need it for retrieve), name, description, status, updated_at. Set include_ids=True for the raw knowledge_base_arn.

#MistralToolSet

Python 
from maivn_tools import MistralToolSet

connector = MistralToolSet(api_key=secrets["MISTRAL_API_KEY"])

Tools: chat_completion, embeddings, fim_completion, list_models.

Agent-ready behavior

  • list_models (default limit 25): summary with model_ref, model_name, owned_by, created. model_name is kept because it is also the identifier callers pass to chat_completion. Set include_ids=True for the raw id.

#CohereToolSet

Python 
from maivn_tools import CohereToolSet

connector = CohereToolSet(api_key=secrets["COHERE_API_KEY"])

Tools: chat, embed, rerank, list_models.

Agent-ready behavior

  • list_models (default limit 20): summary with model_ref, model_name, endpoints, context_length. model_name is the identifier callers pass to chat / embed. Set include_ids=True for the raw name.

#HuggingFaceToolSet

Python 
from maivn_tools import HuggingFaceToolSet

connector = HuggingFaceToolSet(token=secrets["HF_TOKEN"])

Tools: list_models, get_model, list_datasets, run_inference, whoami.

Agent-ready behavior

  • list_models (default limit 25): summary with model_ref, model_name (the "author/name" string callers pass to run_inference), pipeline_tag, downloads, likes, last_modified. Set include_ids=True for the raw id.
  • list_datasets (default limit 25): summary with dataset_ref, dataset_name, tags, downloads, last_modified. Set include_ids=True for the raw id.
Python 
hits = connector.list_models(search="bert", max_results=10)

#ReplicateToolSet

Python 
from maivn_tools import ReplicateToolSet

connector = ReplicateToolSet(api_token=secrets["REPLICATE_TOKEN"])

Tools: list_models, get_model, create_prediction, get_prediction, cancel_prediction (destructive), list_predictions, list_deployments, predict_deployment.

Agent-ready behavior

  • list_models (default limit 25): summary with model_ref, model_name ("owner/name"), owner, name, description, run_count, visibility. Set include_ids=True to surface latest_version_id (needed for pinned-version create_prediction).
  • list_predictions (default limit 25): summary with prediction_ref, version, status, created_at, completed_at. Raw prediction IDs hidden; opt in with include_ids=True (needed for get_prediction / cancel_prediction).
  • list_deployments (default limit 25): summary with deployment_ref, owner, name, deployment_path, current_release_version. Opt in for the raw id.
  • Tolerant inputs: get_prediction and cancel_prediction accept a raw prediction ID, a prediction dict returned by create_prediction / list_predictions, or the {"results": [...]} list payload.
  • Destructive: cancel_prediction. Confirm before invoking.

#StabilityToolSet

Python 
from maivn_tools import StabilityToolSet

connector = StabilityToolSet(api_key=secrets["STABILITY_KEY"])

Tools: get_account, get_balance, generate_image, upscale, edit_inpaint.

Agent-ready behavior

  • No list tools; the only read tools (get_account, get_balance) return small scalar responses already useful as-is.

#ElevenLabsToolSet

Python 
from maivn_tools import ElevenLabsToolSet

connector = ElevenLabsToolSet(api_key=secrets["ELEVENLABS_KEY"])

Tools: text_to_speech, speech_to_text, list_voices, get_voice, list_models, get_user.

Agent-ready behavior

  • list_voices (default limit 25): summary with voice_ref, name, voice_id (kept because callers need it for text_to_speech / get_voice), category, labels, description. Set include_ids=True for a redundant id field.
  • list_models (default limit 25): summary with model_ref, model_name, name, languages, can_do_text_to_speech, can_use_style. Set include_ids=True for the raw model_id.

#DeepgramToolSet

Python 
from maivn_tools import DeepgramToolSet

connector = DeepgramToolSet(api_key=secrets["DEEPGRAM_KEY"])

Tools: transcribe_url, transcribe_bytes, speak, list_projects.

Agent-ready behavior

  • list_projects (default limit 25): summary with project_ref, name, company. Raw project_id is hidden by default; opt in with include_ids=True when a follow-up Deepgram tool needs the raw ID.

#AssemblyAIToolSet

Python 
from maivn_tools import AssemblyAIToolSet

connector = AssemblyAIToolSet(api_key=secrets["ASSEMBLYAI_KEY"])

Tools: upload, submit_transcript, get_transcript, list_transcripts, delete_transcript (destructive), lemur_task.

Agent-ready behavior

  • list_transcripts (default limit 10): summary with transcript_ref, status, audio_url, created, completed. Raw transcript IDs hidden; opt in with include_ids=True (needed for get_transcript, delete_transcript, lemur_task).
  • Tolerant inputs: get_transcript, delete_transcript, and lemur_task accept a raw transcript ID, a transcript dict, or the {"transcripts": [...]} list payload. lemur_task accepts a list of any of the above.
  • Destructive: delete_transcript. Confirm before invoking.
Python 
recent = connector.list_transcripts(include_ids=True)
connector.delete_transcript(recent["transcripts"][0])

#TavilyToolSet

Python 
from maivn_tools import TavilyToolSet

connector = TavilyToolSet(api_key=secrets["TAVILY_KEY"])

Tools: search, extract, crawl.

Agent-ready behavior

  • search is a search API, not a list tool: it returns the raw Tavily response (results[*] with title, url, content) which is already agent-friendly. max_results is capped at 10 (the actual cap is reported back on the response when the requested value exceeded the cap).

#BraveSearchToolSet / SerpAPIToolSet

Two web-search providers with similar surfaces:

Python 
from maivn_tools import BraveSearchToolSet, SerpAPIToolSet

brave = BraveSearchToolSet(api_key=secrets["BRAVE_KEY"])
serp  = SerpAPIToolSet(api_key=secrets["SERPAPI_KEY"])

Each exposes web_search. Brave also has news_search, image_search, video_search and adds suggest. SerpAPI adds location_search and get_account.

The Bing Web Search API was retired by Microsoft on 2025-08-11 and is no longer available; the connector was removed. Use Brave/SerpAPI, or Grounding with Bing Search in the Azure AI Agents Service.

Agent-ready behavior

  • All search tools return raw provider responses (Brave web.results[*], SerpAPI organic_results[*]) -- the per-hit fields (title/name, url/link, snippet/description/content) are already human-readable, so they are kept.
  • count (Brave) and num (SerpAPI) cap results to keep agent runs fast; values above each provider's max are clamped.