agentseed

Minimal pluggable LLM agent server: FastAPI + Pydantic AI with swappable providers (AWS Bedrock, OpenAI, Anthropic, Groq, Ollama) behind one async interface.

Quick Start

uv sync

# Configure a provider (bedrock is the default)
export AWS_PROFILE=my-profile          # Bedrock (SSO)
export AWS_REGION=ap-southeast-2       # Bedrock
aws sso login --profile my-profile
# export CHAT_SERVICE=openai|anthropic|groq|ollama   # override provider
export OPENAI_API_KEY=sk-...           # OpenAI
export ANTHROPIC_API_KEY=sk-ant-...    # Anthropic
export GROQ_API_KEY=gsk_...            # Groq

uv run agentseed
# http://localhost:8000

uv run agentseed check             # which providers are live?

Endpoints

• GET / — chat UI (agentseed/index.html)
• GET /health — liveness
• GET /providers/status[?service=&all_models=] — live-probe providers (a few tokens each)
• GET /config — default provider/model, availability, and all selectable models pairs
• POST /agent/chat[?service=&model=] — native Pydantic AI message history (ModelMessage JSON arrays); optional query params pick a pair from models.json

Frontend

The chat UI is a single file, agentseed/index.html, based on the one-page app template from onepageapp. FastAPI serves it at /. It loads Vue 3, Bootstrap, marked and pug from CDNs, so there is no build step: edit the file and reload the page.

It keeps the conversation in localStorage as Pydantic AI ModelMessage JSON, and has a provider/model dropdown fed by /config. Providers that failed the startup probe are shown disabled, with the error on hover.

Layout

• agentseed/providers.py — Pydantic AI model/embedding builders, CHAT_SERVICE selection (default bedrock), and get_aws_config
• agentseed/server.py — FastAPI app (/, /health, /config, /agent/chat) and the Pydantic AI agent builder
• agentseed/cli.py — agentseed (serve, default) and agentseed check (cyclopts + uvicorn)
• agentseed/logger.py — Rich logging setup
• agentseed/models.json — selectable chat and embedding models
• agentseed/index.html — single-file chat UI (Vue 3 + Bootstrap via CDN)
• tests/ — offline tests for the chat endpoint

Checking providers

agentseed check sends each provider's default model a one-word prompt (capped at 16 output tokens) and prints a table of live/failed providers with latency and the error. It exits non-zero if any probe fails.

uv run agentseed check                 # default model of every provider
uv run agentseed check anthropic groq  # only these providers
uv run agentseed check --all           # every model in models.json

The same probe is available at GET /providers/status.

Tests

uv run pytest