ferrant 0.1.18

Fast, lightweight AI agents powered by a Rust core

ferrant Python wrapper

This directory is an optional, thin Python binding around the Rust ferrant crate. The agent loop, providers, MCP, graph scheduler, persistence, streaming, and retrieval remain implemented in Rust. Installing this package does not change the Rust crate or its examples.

pip install ferrant

Then import it directly:

import asyncio
import os
from ferrant import Agent

async def main():
    agent = Agent.openai("gpt-5-nano", os.environ["OPENAI_API_KEY"])
    print(await agent.run("Explain Rust ownership in one paragraph."))

asyncio.run(main())

For local wrapper development, use maturin develop --release from this directory. End users do not need Maturin or a Rust toolchain when installing a prebuilt wheel.

pip install ferrant also installs the ferrant command and its local FastAPI/Uvicorn runtime dependencies. Run ferrant init to create a function-based Python agent and ferrant run to host it locally. Configure server: https://deploy.example.com and environment: <environment-id> in deploy.yml, create a token in the web console, run ferrant login --email [e-mail] --token fdt_..., then run ferrant deploy to deploy through a Ferrant deployment server. The CLI does not accept account passwords. For CI, set a scoped deploy token in FERRANT_TOKEN. See the repository's deployment documentation for the handler contract.

The first authenticated deploy automatically creates or reuses the manifest's named project when environment still contains the generated placeholder. It selects the development environment and saves its ID back to deploy.yml.

The wheel uses PyO3's stable ABI for Python 3.9+. Rust futures are exposed as normal asyncio awaitables. Python custom-tool and graph-node callbacks are synchronous by design; keep expensive execution in Rust tools, MCP servers, or model calls.

See examples/ for Python counterparts of every top-level Rust example and matched advanced workflow examples.

Focused Python examples

Run these from the python-wrapper/ directory after installing the package:

python examples/streaming.py
python examples/memory.py
python examples/rag.py
DOCUMENT_PATH=invoice.pdf python examples/document_extraction.py
DOCUMENT_PATH=invoice.pdf python examples/rag_document_extraction.py

• streaming.py prints content_delta events as the model generates them.
• memory.py stores a session in .ferrant/sessions and recalls a prior turn.
• rag.py persists a local hybrid vector index, retrieves relevant documents, and provides the matches as grounded agent context.
• document_extraction.py base64-encodes a local PDF and asks OpenAI to extract invoice fields. Set DOCUMENT_PATH to a PDF and OPENAI_API_KEY before running it.
• rag_document_extraction.py transcribes a local PDF with OpenAI, indexes the text locally, retrieves invoice-relevant passages, and returns schema-validated extraction results. Set DOCUMENT_PATH to a PDF.

The exposed surface covers OpenAI-compatible and Anthropic agents, Python tools, MCP-discovered tools, coordinator teams, multimodal input/output, streaming callbacks, schema-validated output, persistent retrieval, and durable workflow graphs with routes, parallel joins, retry/timeout policies, interrupts, resume, and recovery.