The safe runtime layer for enterprise AI agents.
SandBase helps teams move agents from demos to production with runtime infrastructure for sessions, tools, approvals, sandboxed execution, memory, audit trails, replay, and operational visibility.
Website · Docs · Blog · Discord · X · LinkedIn
Modern agents are no longer just chat interfaces. They call tools, run code, inspect files, browse the web, trigger workflows, and act on behalf of users.
That creates a new infrastructure problem:
- where does generated code run?
- which tools can the agent call?
- who is the agent acting for?
- what happens before a sensitive action runs?
- how do teams observe, replay, and debug agent behavior?
SandBase is built around those runtime questions. The open-source path starts with local-first managed agents; the hosted SandBase platform adds managed sandboxes, team controls, observability, connectors, and production support.
| Area | Focus |
|---|---|
| Agent runtime | Sessions, event logs, replay, memory, and resumable agent execution |
| Tool governance | MCP/tools, permissions, approval patterns, and action policies |
| Sandboxed execution | Safer code, shell, file, browser, and workspace operations |
| Multi-executor support | Runtime patterns that can work across Claude, OpenAI, local models, and custom executors |
| Observability | Logs, audit trails, status, and operational surfaces for agent runs |
| Open resources | Ecosystem maps, cookbooks, labs, and growth playbooks for builders |
| Project | What it is for |
|---|---|
| SandBase CLI | One-command MCP onboarding for Cursor, Claude Code, Codex, Windsurf, Gemini CLI, OpenCode, and other agents, with access to 2,000+ tools and 200+ AI models. |
| managed-agents | Open-source, local-first managed-agent runtime with a Console, Claude Managed Agents-style resource APIs, skills, files, credential vaults, memory stores, environments, and resumable session events. |
| awesome-native-agent-platforms | A curated list of infrastructure, runtimes, sandboxes, browsers, model routers, and protocols for building production AI agents. |
| sandbase-lab-sitecheck | "Can AI Get It?", a SandBase-powered website AI personality test where an agent visits a site, scores it, writes feedback, assigns personality tags, and generates a shareable card. |
| awesome-agent-runtime | A 500-project landscape of agent runtimes, sandboxes, browser agents, MCP/tool protocols, memory layers, observability, and compute platforms. |
| agent-sandbox-cookbook | Examples, compatibility checks, and field notes for running AI agent tools across sandboxed runtimes. |
| global-ai-cold-start | A public case study on turning SandBase.ai from an invisible early AI infrastructure product into a searchable, developer-facing trust surface. |
- Want to connect your coding agent to tools and models? Install SandBase CLI.
- New to agent infrastructure? Read the SandBase blog.
- Comparing native agent platforms and runtimes? Start with Awesome Native Agent Platforms.
- Mapping the broader runtime ecosystem? Browse Awesome Agent Runtime.
- Exploring local-first enterprise agent runtime? Browse Managed Agents.
- Building agents that run code or call tools? Read the Agent Sandbox Cookbook.
- Want to see an agent judge a website? Try SandBase Lab Sitecheck.
- Following SandBase updates? Join Discord or follow @SandbaseAI.
We care about infrastructure that helps agents act safely and reliably:
- agent runtime and execution boundaries
- MCP servers, tool protocols, and action schemas
- sandboxed compute for code, shell, browser, and file operations
- model gateways and multi-model routing
- evals, tracing, replay, and observability
- authorization, approvals, and pre-action policy checks
- long-running workflows and distributed execution for agents
If you are building in this direction, we would love to learn from you.