open beta · Apache-2.0

Open-source managed agents for real business work.

Run long-horizon AI agents with your own models, tools, data, sandboxes, and governance. AstraBox connects agents to data warehouses, MCP servers, internal APIs, and enterprise workflows — without locking you into a single model provider.

Start freeBook enterprise pilot$npx create-astrabox-agent data-analyst
Run withDeepSeekClaudeOpenAIQwenKimiGLMLlamaBYOK
02 · Core features

The runtime layer for agents that actually finish the work.

OpenAI Frontier and Claude Managed Agents show where the market is going. AstraBox is the model-agnostic, hosted alternative — with an Enterprise Edition you can run in your own VPC.

AGENT

Agent runtime

Reusable agents bundling model, system prompt, skills, tools, MCP servers, memory and permission policies.

ENVIRONMENT

Sandbox environments

Run agents in isolated containers with files, packages, network controls, and persistent workspaces.

CONTEXT

Business context

Connect data warehouses, docs, CRM, internal APIs, SOPs, metrics, and policies — durable institutional memory.

CONNECTORS

MCP-native connectors

Bring MCP connectors built for ChatGPT or Claude, or build your own with the AstraBox Connector SDK.

ARTIFACTS

Artifacts, not chat

Generate reports, spreadsheets, slides, PDFs, SQL results, evidence packs, and audit exports.

EVALS

Evals & observability

Track success rate, cost per task, model performance, tool failures, and human feedback over time.

03 · Architecture

Agent. Environment. Session. Events.

Every run is reproducible. An agent is loaded into an environment; a session captures the events, tool calls, artifacts, costs, and audit trail.

Bring your own models. Bring your own MCP. Bring your own infrastructure.

AstraBox runtime architecture: agent, session, artifacts inside a sandbox environment connected to connectors, business context, and evals.
04 · Quickstart

Two minutes from sign-up to first agent.

Sign in to AstraBox Cloud, drop in a model key, point it at a Postgres database, and run the Data Analyst Agent on a real business question. Need air-gapped? Enterprise Edition self-hosts the same runtime.

  1. Sign in to AstraBox Cloud
    Two-minute setup — no infrastructure to provision.
  2. Bring your own model key
    DeepSeek, Claude, OpenAI, Qwen, Kimi, GLM, or local.
  3. Connect a data source
    Postgres, ClickHouse, BigQuery, Snowflake, or any MCP server.
  4. Run an agent — get an artifact
    Reports, SQL, evidence packs, audit logs. Not just chat.
~/astrabox · zsh
$ astrabox auth login
→signed in as jordan@eastwind.io · workspace eastwind-data
$ astrabox agents create --template data-analyst
→agent agt_8K2x · model: deepseek-v4-flash
$ astrabox run agt_8K2x \
$ "why did east-region GMV drop last week?"
→session RUN-237 · streaming events…
·toolpostgres.list_schemas
·toolpostgres.query (2.4s)
·toolartifact.write report.md
✓ run completed · 14.2s · $0.42 · 94 events

Build your first managed business agent today.

Spin up your first managed agent in two minutes on AstraBox Cloud. Self-hosted Enterprise Edition is available for regulated workloads.

Start free on CloudBook enterprise pilotView on GitHub