spinekit 0.2.0

A lightweight, modular, protocol-native runtime for production AI agents.

Spine

A lightweight, modular, protocol-native runtime for production AI agents.

Documentation · Quickstart · Cookbook · Contributing

---

Spine is the kernel for AI agents — a tiny, load-bearing runtime that everything else plugs into. Where monolithic frameworks bundle heavy abstractions, Spine ships a small core and pushes every feature into opt-in middleware, swappable backends, and protocol adapters.

from spine_core import Agent

agent = Agent("openai:gpt-4o-mini")
print((await agent.run("say hello")).answer)

Three guarantees

• No hidden prompts — every model call consumes inspectable, typed Message objects.
• No runaway loops — guards are enforced inside the kernel, every iteration.
• Deterministic replay — any run can be recorded and replayed step-for-step.

Install

One package, lean core, opt-in extras:

pip install spinekit                  # kernel only (pydantic + anyio)
pip install "spinekit[openai]"        # + OpenAI
pip install "spinekit[anthropic,redis,cli]"   # pick parts
pip install "spinekit[all]"           # everything

Import name stays spine_core. See the install guide.

What's in the box

Module	Extra	What
spine_core	—	the kernel — loop, state, guards, tracer, protocols, HITL, streaming
spine_middleware	—	retry, fallback, guardrails, cache, memory recall, sandbox, replay, …
spine_backends	redis / postgres	checkpoints (SQLite always; Redis/Postgres) + memory (vector/buffer/pgvector)
spine_providers	openai / anthropic	OpenAI + Anthropic (and any OpenAI-compatible endpoint)
spine_mcp · spine_a2a · spine_otel	mcp · a2a · otel	MCP tools · remote agents · OpenTelemetry
spine_eval	eval	dataset + scorers + Cost/Latency/Efficacy/Reliability report
spine_orchestration	—	sequential / supervisor / handoff
spine_cli	cli	init / run / chat / dev / trace / eval / doctor / plugin

Develop

uv sync                      # install workspace + dev deps
uv run pytest                # tests
uv run ruff check .          # lint
uv run mypy                  # strict type-check
uv run mkdocs serve          # preview docs at http://127.0.0.1:8000

See the Contributing guide to build your own middleware, provider, or backend and publish it as a plugin.

License

MIT © 2026 Research Analytics Solutions