Private Preview · v0.9

Scheduled agents, built on Claude.

Run long-lived agents on top of the Claude API with persistent memory, scheduled triggers, and observability built in. Bring your own tools, your own prompts, and your own data retention policy.

Request access Read the quickstart

Typed tool-use, out of the box

Declare tools with JSON Schema, stream results back to the agent, and branch on whatever the model decides. No orchestrator library to learn.

Durable state

Agents resume from the last event on restart. Memory is versioned; rollback is a single API call.

Scheduled triggers

Cron, webhook, and queue triggers drive runs. Each run is its own immutable trace, queryable from the SDK.

Provenance you can audit

Every tool call, model response, and state transition is signed. Export full traces in OpenTelemetry format.

Quick start

Install the SDK and run your first agent against a message endpoint.

pip install anthropic-sdk-python

from anthropic_agent_runtime import Agent

agent = Agent.from_env("prod")
resp = agent.run(
    "v1/agents/triage/run",
    input={"ticket_id": "T-4721"},
)
print(resp.summary)

Endpoints are gated by API key — x-api-key + anthropic-version headers required on every request. See the auth docs for rotation and scope rules.