Skip to main content
open source · self-hosted · apache-2.0

Run agent engines on your infrastructure.

AstraBox runs existing agent engines in isolated sandboxes. Your team gets a web console and APIs for live output, tool approvals, and saved session history. Deploy with Docker or Kubernetes.

Start locallyExplore capabilities
Model access through LiteLLMAnthropic·DeepSeek·OpenAI-compatible·Gemini·Ollama
build and start AstraBox locally
docker build -t astrabox/server:latest -f containers/server/Dockerfile .
docker build -t astrabox/sandbox-claude-code:latest \
  -f containers/sandbox-claude-code/Dockerfile .

export ANTHROPIC_API_KEY=sk-ant-…
scripts/compose.sh up
01 · Capabilities

Run and recover sessions with existing agent engines

Agent engines

Use the engines you already know

Each installed Agent engine keeps its native reasoning and tool-use behavior. The live Environment schema reports the choices available in this deployment; AstraBox starts the selected engine, passes tasks and events, and records sessions. Configure an Agent →

Sandbox

Isolated workspaces by default

Conversation tenancy gives each Session its own sandbox; supported Agent tenancy can share a sandbox across Sessions. OpenSandbox creates and reclaims them on Docker or Kubernetes, and prewarmed pools can reduce startup time. Isolation strength depends on the configured container runtime, such as Docker, gVisor, or Kata. Explore sandbox support →

Persistence

Save work when a sandbox pauses

With a snapshot-capable Kubernetes setup, pausing saves the filesystem and releases compute. Resuming starts new processes with the saved files; AstraBox stores session history separately. Run a Session →

Models

Operate models through LiteLLM

Use the bundled LiteLLM gateway or connect one your team operates. Bundled installs open LiteLLM's own interface from the AstraBox admin console to manage model routes, budgets, and providers, and inspect gateway-wide spend and request logs. External links appear only when their browser address is configured. Session behavior depends on the engine and model you choose. See the gateway guide →

Security

Protect model and tool credentials

Remote MCP services connect directly from the Agent engine. A compatible sandbox backend adds OAuth, bearer, or API-key credentials only to matching outbound requests. Model credentials use the same protection by default. Manage credentials →

Messaging platforms

Put an Agent where conversations already happen

Connect an Agent to Telegram, Discord, Slack, Feishu, Lark, DingTalk, QQ, KOOK, Zulip, LINE, WhatsApp, WeChat, WeCom, Email, Matrix, or a Satori Protocol server. Configure the bot and its write-only credentials in AstraBox. The official platform adapters are included in the server image. Explore messaging platforms →

02 · Next steps

Choose what you want to do next

Local evaluation

Try AstraBox on one machine

Build the server and sandbox images, start the reference Docker Compose setup, and open the web console. Follow the quickstart →

Team deployment

Add Kubernetes, shared state, and SSO

Move sandbox workloads to Kubernetes, configure a shared database, and use the bundled Casdoor or your existing OIDC provider. When Casdoor is enabled, administrators can open its user, group, MFA, and scoped API-client management directly from AstraBox. Read the deployment guide →

Platform integration

Connect the systems you already operate

Register plugins for agent engines, sandbox backends, model endpoints, data stores, credential stores, or messaging-platform integrations. Explore the developer reference →

03 · Architecture

How an agent session runs

The browser connects to AstraBox, which starts or resumes a sandbox and passes work to the selected agent engine. Live events return to the browser, while session history is stored outside the sandbox for recovery.

web console / APIrequests · live outputsession servicestreaming · approvalsOpenSandboxDocker · Kubernetessession storagehistory · stateagent adaptertasks · eventsagent engineClaude · Hermes · DeepSeekworkspace & toolsfiles · commandsmodel endpointLiteLLM or your gatewayAstraBox servicemanaged sandboxtasks · eventscreate · pause · resumemodel requestssession history
  1. 01web console / APIrequests · live output
  2. 02session servicestreaming · approvals · session storage
  3. 03OpenSandboxDocker · Kubernetes
  4. 04agent engineagent adapter · workspace & tools
  5. 05model endpointLiteLLM or your gateway

Solid lines show live requests and events. The dashed line shows session history stored by AstraBox. On snapshot-capable deployments, pausing preserves sandbox files; resuming starts fresh processes from those files.

04 · Integrations

Included integrations and ways to add your own

A filled dot marks an included implementation. A hollow dot marks an option you can add through a documented Python plugin interface.

Agent engines

  • Installed Agent enginesreported by the live Environment schema
  • Custom engineregister a Python plugin

Sandbox backends

  • Dockermanaged by OpenSandbox
  • Kubernetespools and optional snapshots
  • Custom backendimplement the backend interface

Model endpoints

  • Embedded LiteLLMlocal setup with an isolated database
  • Your own LiteLLMteam gateway with its own database
  • Custom gatewayregister a model endpoint plugin

State & session history

  • PostgreSQLdefault for local and team deployments
  • MongoDBavailable for team deployments
  • Custom data storeregister a repository plugin

Credential storage

  • Encrypted local vaultAES-GCM under a master key
  • AWS KMSbuilt-in envelope encryption for stateless replicas
  • External secret storeadd a secret-store plugin

Messaging platforms

  • JSON Webhookbuilt-in HTTP integration
  • Telegram / Discord / Slack / Feishu / Lark / DingTalk / QQ / KOOK / Zulipbuilt-in bot adapters
  • LINE / WhatsApp / WeChat / WeCom / Email / Matrix / Satoribuilt-in direct and callback adapters
05 · Startup paths

What each startup path does

Running sandbox
Reuse the sandbox and start the next task.
Prewarmed sandbox
Assign a ready sandbox, then start the Session.
New sandbox
Create a sandbox before the Session can start.
Paused sandbox
Restore saved files and start fresh processes.

These paths describe different amounts of startup work. Actual latency depends on the deployment.