0
mtpx
multi tool protocol
pip install mtpx
docs / runtime / sdk

Provider Guides

Provider Guides explains the overview layer of MTPX with practical guidance for building inspectable, tool-using agents.

Pr
Provider Guides
Chapter 016ProvidersOverviewIndex
01Orientation

Provider Guides belongs to the providers track. The page breaks the idea into responsibilities, implementation rules, failure modes, and the signals you should expose in a product UI.

Use this when

Use this when you need to understand how provider guides affects a real MTPX agent before you wire it into an application.

  • What provider guides owns in the runtime.
  • How it connects to planning, tool execution, events, providers, or storage.
  • What to log, test, and expose to users when this layer is active.
  • Common mistakes that make agent systems hard to inspect.
02 / syntax starter
from mtp import Agent
from mtp.providers import Groq

agent = Agent.MTPAgent(
    provider=Groq(model="llama-3.3-70b-versatile"),
    tools=tools,
)
03 / manual

Read the system

The complete source manual—syntax, examples, linked references, edge cases, and implementation notes.

Documentation index ↗
01Common Rules

This file documents the provider adapters exposed by MTP and how they should be used from Agent OS, the TUI, or regular SDK code.

Common Rules

All providers are created explicitly and passed into Agent or MTPAgent.

pythonfrom mtp import Agent
from mtp.providers import Groq

tools = Agent.ToolRegistry()
provider = Groq(model="llama-3.3-70b-versatile")
agent = Agent.MTPAgent(provider=provider, tools=tools)

Provider adapters convert provider-native tool calls into MTP ExecutionPlan objects. The runtime, not the provider, executes tools. See TOOL_CALL_SYNTAX.md for exact tool-call syntax.

Normal chat should use autoresearch=False. Autoresearch is persistent mode: direct model text is progress, and completion is expected through agent.terminate.

02Mock

Mock

  • Alias: MockPlannerProvider
  • SDK package: none
  • API key: none
  • Default model label in Agent OS: simple-planner
  • Best for: deterministic tests and local demos

The mock provider is not a general LLM. It only returns deterministic plans/text for tests.

03Groq

Groq

  • Alias: Groq
  • Class: GroqToolCallingProvider
  • Extra: pip install "mtpx[groq]"
  • Env var: GROQ_API_KEY
  • Default model: llama-3.3-70b-versatile
  • Tool calling: native Groq tool calls only
  • Streaming: finalize streaming with usage capture when supported

Recommended when you want fast cloud inference with OpenAI-style tool schemas.

04OpenAI

OpenAI

  • Alias: OpenAI
  • Class: OpenAIToolCallingProvider
  • Extra: pip install "mtpx[openai]"
  • Env var: OPENAI_API_KEY
  • Default model: gpt-4o
  • Tool calling: native OpenAI function/tool calls
  • Streaming: finalize fallback unless the selected adapter path exposes native streaming

Use models that support tool calling. Keep tool arguments strict JSON.

05OpenRouter

OpenRouter

  • Alias: OpenRouter
  • Class: OpenRouterToolCallingProvider
  • Extra: pip install "mtpx[openrouter]"
  • Env var: OPENROUTER_API_KEY
  • Default model: qwen/qwen-2.5-72b-instruct
  • Tool calling: OpenAI-compatible tool calls, model-dependent

OpenRouter routes to many model families. Tool reliability depends on the selected model. Prefer models advertised as tool/function-call capable.

06Anthropic

Anthropic

  • Alias: Anthropic
  • TUI name: claude
  • Class: AnthropicToolCallingProvider
  • Extra: pip install "mtpx[anthropic]"
  • Env var: ANTHROPIC_API_KEY
  • Default model: claude-3-5-sonnet-20241022
  • Tool calling: Anthropic tool-use blocks

Anthropic uses a different native message shape internally, but MTP still exposes the same event stream and ExecutionPlan runtime semantics.

07Gemini

Gemini

  • Alias: Gemini
  • Class: GeminiToolCallingProvider
  • Extra: pip install "mtpx[gemini]"
  • Env var: GEMINI_API_KEY
  • Default model: gemini-2.0-flash-exp
  • Tool calling: Gemini function calls
  • Multimodal: supports provider-specific media paths where capability checks allow them

Gemini usage and reasoning metrics are normalized through shared usage extraction where available.

08Mistral

Mistral

  • Alias: Mistral
  • Class: MistralToolCallingProvider
  • Extra: pip install "mtpx[mistral]"
  • Env var: MISTRAL_API_KEY
  • Default model: mistral-large-latest
  • Tool calling: Mistral tool calls

Use recent Mistral models for better structured tool output.

09Cohere

Cohere

  • Alias: Cohere
  • Class: CohereToolCallingProvider
  • Extra: pip install "mtpx[cohere]"
  • Env var: COHERE_API_KEY
  • Default model: command-r-plus-08-2024
  • Tool calling: Cohere-compatible chat/tool calls

Cohere is useful for command-style reasoning and retrieval-like tasks. Keep schemas small and descriptive.

10SambaNova

SambaNova

  • Alias: SambaNova
  • Class: SambaNovaToolCallingProvider
  • Extra: pip install "mtpx[sambanova]"
  • Env var: SAMBANOVA_API_KEY
  • Default model: Meta-Llama-3.1-405B-Instruct
  • Tool calling: OpenAI-compatible where model supports it

SambaNova model names can change by account/endpoint. Confirm your available model id before use.

11Cerebras

Cerebras

  • Alias: Cerebras
  • Class: CerebrasToolCallingProvider
  • Extra: pip install "mtpx[cerebras]"
  • Env var: CEREBRAS_API_KEY
  • Default model: llama3.1-70b
  • Tool calling: OpenAI-compatible

Cerebras is optimized for fast hosted inference. Tool behavior is best with concise tool descriptions.

12DeepSeek

DeepSeek

  • Alias: DeepSeek
  • Class: DeepSeekToolCallingProvider
  • Extra: pip install "mtpx[deepseek]"
  • Env var: DEEPSEEK_API_KEY
  • Default model: deepseek-chat
  • Tool calling: OpenAI-compatible

Prefer deepseek-chat or another tool-capable endpoint for agent workflows.

13TogetherAI

TogetherAI

  • Alias: TogetherAI
  • Class: TogetherAIToolCallingProvider
  • Extra: pip install "mtpx[togetherai]"
  • Env var: TOGETHER_API_KEY
  • Default model: meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo
  • Tool calling: OpenAI-compatible for supported models

TogetherAI exposes many models. Tool quality varies by model family.

14FireworksAI

FireworksAI

  • Alias: FireworksAI
  • Class: FireworksAIToolCallingProvider
  • Extra: pip install "mtpx[fireworksai]"
  • Env var: FIREWORKS_API_KEY
  • Default model: accounts/fireworks/models/llama-v3p1-70b-instruct
  • Tool calling: OpenAI-compatible for supported models

Use account-qualified model ids when required by Fireworks.

15Xiaomi

Xiaomi

  • Alias: Xiaomi
  • Class: XiaomiToolCallingProvider
  • Extra: pip install "mtpx[xiaomi]"
  • Env var: MIMO_API_KEY
  • Default model: mimo-v2.5-pro
  • Default base URL: https://token-plan-ams.xiaomimimo.com/v1
  • Tool calling: native OpenAI-compatible calls only
  • Reasoning: supports Xiaomi reasoning/thinking metadata where the model returns it

The adapter automatically manages thinking mode for planning/finalization and disables tool calls after a tool round where required by the endpoint.

16Ollama

Ollama

  • Alias: Ollama
  • Class: OllamaToolCallingProvider
  • Extra: pip install "mtpx[ollama]"
  • Env var: optional OLLAMA_API_KEY for secured hosts
  • Default host: http://localhost:11434
  • Default model: llama3.2:3b
  • Tool calling: native Ollama chat tool calls when supported by the local model

Install Ollama, pull a model, and verify the server:

bashollama pull llama3.2:3b
ollama list
17LM Studio

LM Studio

  • Alias: LMStudio
  • Class: LMStudioToolCallingProvider
  • Extra: pip install "mtpx[lmstudio]"
  • API key: not required for local use; a dummy value is accepted by the local OpenAI-compatible server
  • Default base URL: http://127.0.0.1:1234/v1
  • Default model: qwen3
  • Tool calling: OpenAI-compatible if the loaded model supports tools

Start the LM Studio local server and load a tool-capable model before launching Agent OS.

18Troubleshooting

Troubleshooting

  • Repeated answers in Agent OS: ensure Persistent Autoresearch Mode is off for ordinary chat.
  • Raw <tool_call> in the answer: use a tool-capable model/provider and configure native tool calls. MTP does not parse tool calls from assistant text.
  • Tool validation failure: compare the model arguments against TOOL_CALL_SYNTAX.md.
  • Missing SDK: install the matching extra, for example pip install "mtpx[groq]".
  • Missing API key: set the documented env var or enter the key in Agent OS.
01

Read

Understand where Provider Guides sits in the agent loop before adding abstractions.

02

Wire

Connect the smallest useful provider, registry, store, or event stream first.

03

Observe

Expose events, logs, results, and failure states while the runtime is still moving.

04

Harden

Add policy, tests, retries, and audit traces after the behavior is visible.

Next docPublishing MTPX To PyPI