Overview
The open-source, self-hosted alternative to Claude Managed Agents.
Mango runs managed agents as an API service in your own deployment. Define an Agent, start a Session, and send it a task. Mango drives the model-and-tool loop, keeps the conversation, and coordinates execution and recovery.
Your application owns the user experience and business tools. Mango manages the running Session; you operate its API, state, and workers. Connect through the HTTP API or the Go, Python, and TypeScript/JavaScript SDKs.
Quickstart
Start Mango locally and complete your first Session without model credentials.
Core concepts
Understand Agents, Sessions, Events, and where tools run.
Alpha
Mango's API is still evolving. Production deployment is not yet supported; review capabilities and limits before choosing a workflow.
What Mango manages
- Agent execution: model calls, context management, tool coordination, and the loop that continues until a turn finishes or needs input.
- Persistent Sessions: accepted input, conversation history, and action waits survive API and orchestration-worker restarts. Stream progress, send follow-ups, interrupt work, and respond to approval requests through the API.
- Tools and context: run shell and file tools in your Docker worker, connect application tools or MCP servers, and reuse versioned Skills and Memory.
- Teams and scheduled work: delegate to persistent specialist Threads, share a Session's budget and workspace, schedule new Sessions, and receive lifecycle Webhooks.
Choose your next step
SDKs
Install the current source client and use typed resource services.
Guides
Configure a model, run a Docker worker, or coordinate multiple agents.
Examples
Explore a terminal client, approval gate, specialist team, and coding workflow.
API reference
Find request shapes, responses, pagination, and event-stream behavior.
Run it yourself
Mango's API, orchestration, and state run in your deployment. The default
self_hosted Environment sends shell and file work to an operator-run worker.
Model calls use your configured endpoint; Web Search/Fetch and MCP have their
own execution paths. See core concepts for the split.
The local stack includes PostgreSQL, Temporal, NATS, and S3-compatible object storage. Read Deployment for configuration and Architecture for how accepted work survives failures.
Mango is an independent project with its own API and SDKs. Design provenance explains the concepts adopted from Claude Managed Agents and the choices made for Mango's self-hosted runtime.