start
Installation
mantis-agent-sdk is a normal Python package. No build tools, no native
extensions, no required system dependencies.
Requirements
- Python ≥ 3.11
- One of: a hosted backend you can reach over HTTPS (OpenAI, Together, Fireworks, Groq, OpenRouter, …), a local Ollama install, or a local llama.cpp install.
Install
pip install mantis-agent-sdkThat covers the hosted-backends path: query, ClaudeSDKClient, tools,
MCP, sessions, sub-agents, hooks, budget — all of it.
Optional extras
The package has no required optional groups, but a few extras pull in provider-specific clients for convenience:
| Extra | What it adds |
|---|---|
mantis-agent-sdk[bedrock] |
boto3 for the AWS Bedrock runtime adapter |
mantis-agent-sdk[all] |
Every backend's optional client (currently just bedrock) |
mantis-agent-sdk[dev] |
pytest, pytest-anyio, ruff, respx for tests |
For Anthropic (anthropic_passthrough), OpenAI, Gemini, vLLM, Ollama,
Together, Fireworks, Groq, OpenRouter, Cerebras, llama.cpp, and TGI — no
extras are required. The provider uses httpx directly.
Verify
import mantis_agent
print(mantis_agent.__version__)You should see something like 0.1.0.
CLI
The package installs a single console script:
mantis-agent --helpSubcommands:
mantis-agent setup-local— install Ollama, pull a CPU-friendly model, run a smoke test. Supports Linux, macOS, and Windows.mantis-agent setup-local-llamacpp— same idea, but builds llama.cpp from source and pulls a GGUF model.
See Local setup for details.