replai 0.1.0

Local-first, framework-agnostic debugger for LLM agents — see, replay, and diff what your agent did.

replai

A local-first debugger for LLM agents. See exactly what your agent did — every model call, tool call, and decision — step by step. No account, no cloud, no Docker. Just pip install and look.

⚠️ Early alpha (v0.1). The capture engine and local viewer work today. Replay and run-diffing are next on the roadmap.

Why

When an AI agent does the wrong thing, you're usually staring at a wall of logs trying to reconstruct what happened. Production observability platforms exist, but they're heavy — dashboards, servers, accounts — built for monitoring at scale, not for the moment you're on your laptop going "wait, why did it call that tool?"

replai is the other thing: a debugger for the dev inner loop. Drop it in, run your agent, and get a clickable, step-by-step timeline of everything it did — locally.

Install

pip install "replai[viewer]"

Quickstart

import replai
replai.init()          # auto-captures Anthropic & OpenAI calls

# ... run your agent exactly as you normally would ...

Then open the viewer:

replai ui

Want to annotate your own steps?

with replai.run("my-agent"):
    with replai.span("retrieve", type="tool_call") as s:
        s.output = my_retriever(query)

Or decorate functions and tools:

@replai.tool
def web_search(query): ...

@replai.trace
def plan(goal): ...

Try it with no API keys

python example.py
replai ui

How it works

• Auto-instrumentation wraps the Anthropic / OpenAI clients, so calls are captured with zero code changes.
• @replai.trace / @replai.tool / replai.span() annotate your own functions and tool calls. Spans nest automatically.
• Everything is stored in a local SQLite file (~/.replai/replai.db). Nothing leaves your machine.
• A small FastAPI viewer renders each run as a step-by-step timeline.

Roadmap

• Capture engine (LLM + tool + function spans, sync & async)
• Local timeline viewer
• Replay — step through a run; re-run from any step
• Diff — compare two runs, highlight where they diverged
• Framework adapters (LangChain, LlamaIndex, …)
• MCP tool-call capture
• OpenTelemetry GenAI export

License

MIT