AgentMuru

Build agents you can see, steer, and trust.

AgentMuru is a Python runtime and browser Workspace for observable, human-governed AI applications. It keeps model turns, tool calls, approvals, events, sessions, artifacts, workflows, and traces inside one explicit application model.

Run the local MVP

python -m pip install agentmuru==0.3.0
muru doctor
muru init my-agent --name "My Agent"
cd my-agent
muru run app:application

Open http://127.0.0.1:8000. The default starter uses FakeModel, so it runs without a credential or network call.

Use an official model provider

muru init my-openai-agent --provider openai
muru init my-anthropic-agent --provider anthropic
muru init my-google-agent --provider google

AgentMuru 0.3 includes official SDK adapters for OpenAI Responses, Anthropic Messages, and Google Gen AI. Each adapter streams into the same normalized text, tool-call, completion, usage, and failure events. API keys stay in provider-standard environment variables.

Add durable local persistence

from agentmuru import Agent, Application, FakeModel, Runtime, SQLitePersistence

persistence = SQLitePersistence("agentmuru.db")
application = Application(
    agent=Agent(
        name="support",
        instructions="Answer support questions using approved tools.",
        model=FakeModel.responses("AgentMuru is ready."),
    ),
    session_store=persistence.sessions,
    artifact_store=persistence.artifacts,
)
runtime = Runtime(application, approvals=persistence.approval_service())

SQLitePersistence stores sessions, complete assistant tool calls, runs, ordered events, artifacts, approvals, and idempotency keys in one file. Operate one active AgentMuru runtime process per SQLite file.

MVP scope

• Python 3.10 through 3.13.
• Typed agents and tools with deny-by-default permissions.
• Human approval for sensitive tool execution.
• Official OpenAI, Anthropic, and Google model adapters.
• In-memory or SQLite session, approval, and artifact stores.
• FastAPI HTTP and WebSocket transport with the bundled React Workspace.
• Explicit workflows, handoffs, trace spans, usage, replay, cancellation, and recovery.
• Optional Databricks adapters outside the core dependency set.

The Go-native adaptive runtime remains an experimental Labs track and is not part of the PyPI 0.3.0 package.

Develop

git clone https://github.com/AjayAJ2000/AgentMuru.git
cd AgentMuru
python -m pip install -e ".[dev,docs,providers]"
python -m pytest -q
python -m mkdocs build --strict

Read the full documentation at https://ajayaj2000.github.io/AgentMuru/.

AgentMuru is available under the MIT License.