The open-source, self-hosted alternative to Claude Managed Agents.

Documentation · Quickstart · API reference · Examples

Define an Agent, start a Session, and send it work through an API. Mango runs the model-and-tool loop, persists the conversation, and coordinates execution and recovery. Build your application with Go, Python, TypeScript/JavaScript, or the HTTP API.

Why Mango?

• Delegate execution. Configure instructions and tools, then let Mango drive the agent loop, manage context, and coordinate tool calls.
• Own the infrastructure. Self-host the API, orchestration, state, and tool workers. The first-party Docker launcher runs shell and file operations on your infrastructure; model calls use your configured endpoint.
• Keep work running. Persist accepted input, conversation history, and action waits across API and orchestration-worker restarts.
• Stay in control. Stream responses, send follow-ups, interrupt a turn, or wait for a person or your application to return a tool result.
• Build beyond one agent. Coordinate persistent specialists, reuse Skills and Memory, schedule Sessions, and subscribe to lifecycle Webhooks.

Important

Mango is alpha. APIs may change, and production deployment is not yet supported. See capabilities and limits for the supported scope of each workflow.

Quickstart

You need Git, Docker with Compose, and make. Start the local stack with its built-in offline model; no model credentials are needed:

git clone https://github.com/yanpgwang/mango.git
cd mango
export MANGO_API_KEY=sk-mango-local-development
MANGO_MODEL_BASE_URL= MANGO_MODEL_API_KEY= MANGO_MODEL_ID= \
  docker compose -f deployments/local/compose.yaml up -d --build
make local-health

With curl and jq installed, run a complete first Session:

bash examples/sdk-quickstart.sh

The example creates an Environment, Agent, and Session, sends a message, reads the persisted reply, and cleans up its resources. Look for Quickstart completed. This is a text-only walkthrough; running shell or file tools requires a separate Environment worker. The local key above is for development only.

Prefer an SDK? The Quickstart walks through the same flow in TypeScript, Python, Go, and HTTP. The current resource-based SDKs must be installed from source; published alpha 1 packages use an earlier interface. Follow the SDK installation guide.

Stop the stack while keeping its data:

make local-down

Explore Mango

I want to…	Start here
Understand Agents, Sessions, and workers	Core concepts
Connect a model endpoint	Model configuration
Run sandboxed shell and file tools	Docker worker guide
Add human input or coordinate a team	Runnable examples
Inspect Sessions in a terminal	Terminal UI
Integrate an API operation	API reference
Understand deployment and recovery	Deployment · Architecture

Mango is an independent project, unaffiliated with Anthropic. Claude Managed Agents is a design reference; Mango owns its API and does not require or proxy a hosted agent service. See design provenance for adopted concepts and intentional differences.

Contributing

Bug reports, documentation fixes, and focused pull requests are welcome. Start with CONTRIBUTING.md for setup and the relevant tests.

make verify       # lint, unit tests, race tests, and vet
make docs-check   # check and build the documentation site

Report vulnerabilities privately through the process in SECURITY.md.

License

Apache License 2.0.