Providers

Agentmd supports six LLM providers: Google, OpenAI, Anthropic, Ollama, Local (self-hosted), and OpenRouter. Any other OpenAI-compatible endpoint works from config alone—give it a provider name and a base_url. Each named provider is an optional dependency—install only what you need.

Supported Providers

Provider	Package	Default Model / Notes
`google`	`langchain-google-genai`	`gemini-2.5-flash`
`openai`	`langchain-openai`	`gpt-4o`
`anthropic`	`langchain-anthropic`	`claude-sonnet-4`
`ollama`	`langchain-ollama`	User-defined
`local`	`langchain-openai`	User-defined (OpenAI-compatible)
`openrouter`	`langchain-openrouter`	Model format `provider/model` (e.g. `anthropic/claude-sonnet-4`)

Installation

Core Installation

pip install agentmd
# or with uv:
uv pip install agentmd

Provider-Specific Installation

# OpenAI
pip install agentmd[openai]

# Anthropic
pip install agentmd[anthropic]

# Ollama
pip install agentmd[ollama]

# OpenRouter
pip install agentmd[openrouter]

# All providers
pip install agentmd[all]

Note: The local provider and any custom OpenAI-compatible endpoint use the same package as OpenAI (langchain-openai).

API Key Setup

Create a .env file in your project root:

GOOGLE_API_KEY=your-key-here
OPENAI_API_KEY=your-key-here
ANTHROPIC_API_KEY=your-key-here
OPENROUTER_API_KEY=your-key-here

Agentmd automatically loads .env at startup via python-dotenv. Keys are read from <PROVIDER>_API_KEY by default (e.g. groq → GROQ_API_KEY). Override with model.api_key_env only when the variable name disagrees with the provider name.

Get Your API Keys

Google: Google AI Studio
OpenAI: OpenAI Platform
Anthropic: Anthropic Console
OpenRouter: OpenRouter Keys

Ollama and loopback local endpoints don't require API keys.

Google

Fast, cost-effective, multimodal—best for high-frequency tasks.

Installation:

pip install agentmd
# Google is included by default

Configuration:

model:
  provider: google
  name: gemini-2.5-flash

Available models: - gemini-2.5-flash — Fast, multimodal (recommended) - gemini-2.5-pro — More capable, larger context - gemini-2.0-flash-exp — Experimental

Example:

---
name: quick-summarizer
model:
  provider: google
  name: gemini-2.5-flash
trigger:
  type: manual
settings:
  temperature: 0.3
---

Summarize the file at `input.txt` and save to `summary.txt`.

OpenAI

Versatile, excellent reasoning—industry standard for complex tasks.

Installation:

pip install agentmd[openai]

Configuration:

model:
  provider: openai
  name: gpt-4o

Available models: - gpt-4o — Latest flagship - gpt-4o-mini — Faster, cheaper - gpt-4-turbo — Previous generation - o1 / o1-mini — Advanced reasoning

Example:

---
name: research-assistant
model:
  provider: openai
  name: gpt-4o
trigger:
  type: manual
settings:
  temperature: 0.7
  max_tokens: 8192
---

Research the topic in `query.txt` and write a comprehensive report.

Anthropic

Long context (200K+ tokens), excellent for detailed analysis.

Installation:

pip install agentmd[anthropic]

Configuration:

model:
  provider: anthropic
  name: claude-sonnet-4

Available models: - claude-sonnet-4 — Balanced performance - claude-opus-4 — Highest capability - claude-haiku-4 — Fast and efficient

Example:

---
name: document-analyzer
model:
  provider: anthropic
  name: claude-sonnet-4
trigger:
  type: manual
settings:
  temperature: 0.5
  max_tokens: 4096
---

Analyze the legal document in `contract.txt` and extract key terms.

Ollama

Local, offline, privacy-first. No API key required.

Setup:

Install Ollama: ollama.com/download
Pull a model: ollama pull llama3.2
Ollama runs at http://localhost:11434 by default

Installation:

pip install agentmd[ollama]

Configuration:

model:
  provider: ollama
  name: llama3.2

For a remote Ollama server, set its native endpoint explicitly. Agentmd sends this URL unchanged (Ollama does not use the OpenAI /v1 prefix):

model:
  provider: ollama
  name: llama3.2
  base_url: http://ollama.internal:11434

Popular models: - llama3.2 — Meta's Llama (3B, 7B, 70B variants) - mistral — Mistral 7B - phi3 — Microsoft Phi-3 (small) - gemma2 — Google Gemma 2

Browse all: ollama.com/library

Example:

---
name: private-summarizer
model:
  provider: ollama
  name: llama3.2
trigger:
  type: manual
settings:
  temperature: 0.7
---

Summarize confidential files without sending data to external APIs.

OpenRouter

One API key for 400+ models across providers. Model names use provider/model format.

Installation:

pip install agentmd[openrouter]

Configuration:

model:
  provider: openrouter
  name: anthropic/claude-sonnet-4

Set OPENROUTER_API_KEY in .env (get a key).

Cost is taken from what OpenRouter reports it actually charged, so max_cost_usd is enforced against a real number. agentmd validate rejects models that cannot call tools, using the capability table bundled with the provider package.

Example:

---
name: openrouter-agent
model:
  provider: openrouter
  name: anthropic/claude-sonnet-4
trigger:
  type: manual
---

Use any OpenRouter-hosted model through a single key.

Local (OpenAI-Compatible)

Self-hosted or third-party OpenAI-compatible endpoints (vLLM, LM Studio, LocalAI).

Installation:

pip install agentmd[openai]

Configuration:

model:
  provider: local
  name: your-model-name
  base_url: http://localhost:8000

The base_url is automatically normalized (e.g., http://localhost:8000 becomes http://localhost:8000/v1).

Common endpoints: - vLLM: http://localhost:8000 - LM Studio: http://localhost:1234 - LocalAI: http://localhost:8080

Example:

---
name: custom-model-agent
model:
  provider: local
  name: my-fine-tuned-model
  base_url: http://localhost:8000
trigger:
  type: manual
---

Use your custom fine-tuned model hosted on vLLM.

Any OpenAI-compatible endpoint

Any provider name plus a base_url is treated as an OpenAI-compatible endpoint—no new code and no release from us. Groq, Together, DeepSeek, xAI, Fireworks, vLLM, and LM Studio all work this way.

Rule: the key is read from <PROVIDER>_API_KEY unless api_key_env says otherwise. An unknown provider name requires base_url; without it, validation fails (typo protection).

Installation:

pip install agentmd[openai]

Example — Groq:

model:
  provider: groq
  name: llama-3.3-70b-versatile
  base_url: https://api.groq.com/openai

# .env
GROQ_API_KEY=gsk-...

base_url is normalized to end in /v1 when missing. If the env var name does not match the provider (e.g. provider gemini but key in GOOGLE_API_KEY), set api_key_env:

model:
  provider: gemini
  name: gemini-2.0-flash
  base_url: https://generativelanguage.googleapis.com/v1beta/openai
  api_key_env: GOOGLE_API_KEY

Common Settings

All providers support:

settings:
  temperature: 0.7      # 0.0 = deterministic, 1.0 = creative
  max_tokens: 4096      # Maximum output tokens
  timeout: 300          # Execution timeout in seconds

Temperature Guide

0.0 — Deterministic (data extraction, structured outputs)
0.3 — Focused (summaries, factual tasks)
0.7 — Balanced (general-purpose, default)
0.9 — Creative (writing, brainstorming)

Switching Providers

Change providers by editing the agent file:

# Before
model:
  provider: openai
  name: gpt-4o

# After
model:
  provider: google
  name: gemini-2.5-flash

Save and run—the next execution uses the new provider.

Troubleshooting

Provider Package Not Installed

ImportError: Provider 'openai' requires langchain-openai.

Solution:

pip install agentmd[openai]

API Key Not Found

Error: OPENAI_API_KEY not found

Solution: 1. Create .env in project root

Add: OPENAI_API_KEY=your-key
Restart the runtime

Scheduled agents whose key is missing are refused at schedule time (error names the variable) instead of registering and failing when the run starts.

Ollama Connection Error

Error: Could not connect to Ollama at http://localhost:11434

Solution: 1. Install Ollama: ollama.com/download

Start Ollama: ollama serve
Pull a model: ollama pull llama3.2

Local Provider 404 Error

Error: 404 Not Found at http://localhost:8000/v1/chat/completions

Solution: - Verify your server is running

Check the correct port (vLLM: 8000, LM Studio: 1234)

Token Limit Exceeded

Error: maximum context length exceeded

Solution: - Reduce max_tokens in settings

Use a model with a larger context window
Split large inputs into smaller chunks