SandBase is an agent-native platform that gives AI agents (like Codex, Claude, and Sandy) a unified runtime for sandboxes, models, MCP servers, skills, long-running sessions, and production workflows through API, CLI, or MCP.
What is SandBase?
SandBase is an agent-native platform for AI agents, offering sandboxes, models, MCP servers, skills, long-running sessions, and production workflows. It takes agent requests as input and produces sandbox environments, model responses, and session outputs. The platform is accessed via API, CLI, or MCP from the hosted service at sandbase.ai.
Key Features
- Agent-native runtime — Execute durable agents with state and traces across sandboxes, models, and MCP servers.
- Multiple access paths — Use the same runtime via REST API (
POST /v1/run), CLI (sandbase session create), or MCP server configuration for agent tool access.
- builder.md context file — Provide a single file to Codex, Claude, Sandy, or any team member that maps which access path to use for each task.
- Long-running sessions — Create and poll sessions for extended agent workflows, with streaming support.
- API key management — Console for creating API keys and managing team members, skills, and MCP publications.
- Live example (Sitecheck) — A real app loop: user submits a URL, SandBase creates a session, reviews the site, saves structured results, and generates a share card.
Who is it for?
- AI developers building production backends for agentic applications.
- Coding assistants like Codex, Claude, or Sandy that need a platform to run sandboxes and execute skills.
- DevOps/CI teams that need CLI-based automation for agent workflows.
- Tool builders creating MCP servers for agents.
What can you do with SandBase?
- App backends: Use the API to create agent sessions that run code, review PRs, or analyze sites.
- Terminal workflows: Use the CLI in CI pipelines or directly in the terminal to manage sandbox sessions.
- Agent tool access: Configure MCP servers to give AI agents (e.g., Claude, Codex) sandbox and model capabilities within conversations.
- Live demos: The Sitecheck lab shows a concrete app loop — user submits a URL, SandBase creates a session, polls events, saves structured result, and generates a share card.
How does SandBase work?
- The developer or agent reads
builder.md to choose the appropriate access path (API/CLI/MCP).
- Using the chosen access layer, they send a request to SandBase's runtime (e.g., create an agent session, run a model).
- SandBase executes the request in a sandboxed environment, streams events back, and persists state and traces for durable agents.
Pricing
Not specified.
FAQ
Is SandBase free?
The page does not specify pricing. You can get an API key from the console – presumably there is a free tier or trial, but no details are provided.
What is builder.md?
It is a context file that tells AI agents (Codex, Claude, Sandy) and developers when to use API, CLI, or MCP for a given task. It serves as a single starting point.
What MCP servers does SandBase support?
SandBase allows configuring MCP servers and publishing skills as MCP servers. The page lists "MCP servers" as one of the runtime services.
Can I run SandBase on my own infrastructure?
The page describes cloud API access; there is no mention of self-hosting. It is offered as a hosted platform at sandbase.ai.
What programming languages are supported?
The example uses Python with the Anthropic SDK, but the API is REST-based so any language can be used.
