Build agents you can see, steer, and trust.

One Python runtime for model calls, governed tools, durable state, and an operator-ready workspace.

Run the quickstart Connect a model

Start locally.
No credential is required for the first run.

python -m pip install agentmuru==0.3.0
muru init my-agent

Choose what you need

Run the MVP

Create a project, send a message, inspect its timeline, and reopen the session.

Follow the quickstart

Use a real model

Choose OpenAI, Anthropic, or Google without changing runtime code.

Configure a provider

Build governed behavior

Turn typed Python functions into tools, require permissions, pause risky work for approval, and replay every decision.

Build a governed tool

One explicit execution path

MessagePersist user intent

ModelStream normalized events

PolicyCheck permission and risk

ToolApprove and execute

TimelineReplay the result

AgentMuru records the assistant message and its complete tool calls before it stores tool results. That ordering keeps provider conversations valid across replay, restart, and a later model turn.

What ships in 0.3

Area	Included
Runtime	Agents, normalized model events, tools, permissions, approvals, cancellation, handoffs
Providers	OpenAI Responses, Anthropic Messages, Google Gen AI, deterministic FakeModel
State	In-memory stores and durable SQLite sessions, events, artifacts, approvals, idempotency
Operator surface	FastAPI HTTP API, WebSocket event stream, bundled browser Workspace
Operations	Health endpoint, trusted hosts, origin controls, trace spans, usage, recovery codes

Clear boundary. The PyPI package is the supported Python-first MVP. Native compilation, adaptive routing, and local-model catalog work live under Labs.

Continue from here

• Understand agents and models
• Persist sessions with SQLite
• Read the Python API reference
• Check current capabilities and limits