BeeOS Managed Agents

BeeOS Managed Agents (BMA) runs long-lived AI agents for you. You define an agent (model, instructions, tools), pick an environment (where its machines come from), and start sessions. BMA runs the agent loop on the server, executes tools on sandboxed or connected machines, and records every step in a durable event log you can read or stream.

BMA is in early access. Every page labels capabilities as Available, Preview (works today, interface or coverage may change) or Roadmap (not built yet). Anything without a label is Available.

Brain, hands, session

BMA keeps three things separate, following the split Anthropic describes for Claude Managed Agents:

Part	What it is	In BMA
Session	The durable record of a conversation: inputs, model output, tool calls and results	An append-only event log in the control plane. It survives machine loss, worker restarts and reconnects.
Brain (harness)	The loop that calls the model, decides which tool to call and manages the context window	The BMA server harness by default. An external brain can run on your own computer through the Executor Protocol (Preview).
Hands (machines)	Where tools run: shell commands, processes, files, a desktop	Sandboxes from a machine pool, created on demand, or your own computer connected with `bmad` or the executor SDK.

Because the session lives outside the machine, a machine can be paused, reclaimed, rebuilt or restored without losing the conversation, and the brain can be replaced without changing how clients read events.

Two APIs, one platform

Surface	Base path	Use it when	Client
Managed Agents–compatible API	`/v1`	You want the Claude Managed Agents object model: agents, environments, sessions, events	The official Anthropic SDKs (Python, TypeScript, Go, …) with `base_url` set to BMA, plus `bma-sdk` for the few BMA extensions
Native REST API	`/api/v1`	You need everything BMA has: orgs and projects, agent configs and profiles, machines and pools, terminals and desktops, artifacts, secrets	`bma-sdk` (Go, TypeScript, Python)

Both read and write the same objects: a session created through /v1 is an agent session in the native API. The compatibility page lists exactly which Claude Managed Agents fields and endpoints BMA serves. A third protocol, the Executor Protocol, connects machines and external brains to BMA. You only need it when you build your own hand or brain, using bma-executor-sdk.

Base URLs and keys

During early access the BMA team gives you an API origin and an organization API key (bma_org_v1_…). In these docs:

Variable	Value
`BMA_BASE_URL`	Your API origin, for example `https://<your-bma-host>`. The Managed Agents API is at `$BMA_BASE_URL/v1`, the native API at `$BMA_BASE_URL/api/v1`.
`BMA_API_KEY`	The organization API key. Send it as `x-api-key` or `Authorization: Bearer`.
`BMA_PROJECT_ID`	The project (`proj_…`) to work in. Send it as `x-bma-project-id`; required when the key can see more than one project.

The SDKs default to https://api.bma.beeos.ai, which is not live yet. Always pass your base URL explicitly.

Start here

Quickstart

Create an agent, an environment and a session, then stream the reply. Python, TypeScript, Go and curl.

Compatibility

What the Claude Managed Agents–compatible API supports, and what it rejects

SDKs

Which SDK to use for what: Anthropic SDKs, bma-sdk, bma-executor-sdk

Executor & brains

Bring your own machine or your own agent loop

Get started

Concepts

Executor & brains

Brain, hands, session

Two APIs, one platform

Base URLs and keys

Start here

Quickstart

Compatibility

SDKs

Executor & brains

​Brain, hands, session

​Two APIs, one platform

​Base URLs and keys

​Start here

Quickstart

Compatibility

SDKs

Executor & brains

Brain, hands, session

Two APIs, one platform

Base URLs and keys

Start here