Overview
coreloop is an open implementation of the Claude Managed Agents API. It runs the agent loop, the tools and the sandbox. You send a message and read the events.
The wire contract is Anthropic's, so the Anthropic SDKs work against coreloop with the base URL changed to https://coreloop.run/anthropic. The difference is where the tokens go. You add your own OpenRouter, OpenAI or Anthropic key and name any model those keys serve. Each turn bills to your account.
Core concepts
| Concept | What it is |
|---|---|
| Agent | The model, the system prompt, the tools and the skills. |
| Environment | Where a session runs. One sandbox per session. |
| Session | One run of an agent in an environment. It holds the transcript. |
| Event | One entry in that transcript, in either direction. |
| Skill | A folder the session mounts, for work the prompt cannot carry. |
How it works
| Step | What happens | |
|---|---|---|
| 1 | Create an agent | Name it, pick a model, write the system prompt, and list the tools. The agent is versioned, and a session pins one version. |
| 2 | Create an environment | An environment names what the tools run on. Sessions share the configuration, and each one runs on its own. |
| 3 | Start a session | A session references one agent and one environment. It keeps running after your request returns. |
| 4 | Send events and read the stream | Every message, tool call and status change is an event. Read them as a page or as server-sent events. |
| 5 | Steer or interrupt | Send a message while the agent works and the agent reads it on the next turn. Send an interrupt to stop the current one. |
What the agent can do
The hosted toolset, agent_toolset_20260401, gives the model a persistent bash shell and a filesystem, in a Linux container of its own: bash, read, write, edit, glob and grep, and the web through web_search and web_fetch. Each tool takes a permission policy, so a tool can stop the turn and ask before it runs.
Soon: MCP servers.
What you can build on it
Any agent. A chat agent and a two-hour autonomous run take the same API and get the same session. The transcript is written whether or not anything is reading it. The run continues after your request returns. A client that drops reads the transcript back and opens a new stream.
The environment decides what the work costs. { type: "cloud" } gives the session a Linux container, as the Managed Agents API does.
Soon: a second environment type that runs the tools in memory, with no container to start.
Compatibility
coreloop tracks the Managed Agents beta of 2026-04-01. One conformance suite drives the Anthropic SDK against coreloop and against Anthropic's own API, and asserts the same contract on each. Where a feature is missing, its page says so. The reference lists every gap.
coreloop is an independent product and is not affiliated with Anthropic.