Installation¶

pip install agentmaker

Requires Python 3.12+.

What the core install includes¶

The base package works out of the box, with no optional extras:

Multi-provider LLM calling over the OpenAI-compatible protocol (openai)
Structured output via Pydantic (pydantic)
JSON Schema validation of tool arguments (jsonschema)
A local vector store, SqliteVecStore (sqlite-vec)
CJK-aware keyword search for the FTS5 index (jieba)
SKILL.md frontmatter parsing for skills (pyyaml)

Optional extras¶

Install only the capabilities you need. Each backend is imported lazily, so an extra you do not install costs nothing at import time.

pip install "agentmaker[anthropic]"        # one extra
pip install "agentmaker[anthropic,rag]"    # several
pip install "agentmaker[all]"              # everything below

Extra	Adds
`anthropic`	Anthropic native protocol adapter
`gemini`	Google Gemini native protocol adapter
`search`	`SearchTool` backends: DuckDuckGo (no key needed), Tavily, Brave, SerpAPI
`rag`	Document loading for RAG: PDF / DOCX / HTML to Markdown
`rerank`	Cohere multilingual reranker
`mcp`	MCP (Model Context Protocol) tool integration
`otel`	OpenTelemetry trace export
`devtools`	Trace Detective: local web UI for diagnosing agent runs

Provider API keys¶

LLMClient reads the API key for a provider from your environment. Set the variable that matches the provider you use, for example:

Provider (argument to `LLMClient`)	Environment variable
`openai`	`OPENAI_API_KEY`
`deepseek`	`DEEPSEEK_API_KEY`
`anthropic`	`ANTHROPIC_API_KEY`
`gemini`	`GEMINI_API_KEY` (or `GOOGLE_API_KEY`)
`dashscope`	`DASHSCOPE_API_KEY`
`moonshot`	`MOONSHOT_API_KEY`
`zhipu`	`ZHIPUAI_API_KEY`
`modelscope`	`MODELSCOPE_API_KEY`

Set the variable in your shell before running:

export DEEPSEEK_API_KEY="sk-..."          # macOS / Linux

$env:DEEPSEEK_API_KEY = "sk-..."          # Windows PowerShell

Prefer a .env file? The variable names inside .env are exactly the ones in the table above — a .env file is just another way to set the same environment variables:

# .env
DEEPSEEK_API_KEY=sk-...

agentmaker deliberately does not read .env itself (loading environment files is an application-level decision), but the standard python-dotenv works as usual:

from dotenv import load_dotenv
load_dotenv()                             # reads .env from the working directory

from agentmaker import Agent, LLMClient
agent = Agent("assistant", LLMClient("deepseek"))

Keep .env out of version control (add it to .gitignore).

Local engines (ollama, vllm, sglang) need no key. See LLM clients & providers for the full provider list and the provider:model syntax.

Installing with uv¶

If you use uv:

uv add agentmaker
uv add "agentmaker[all]"