Agent Plane

AgentPlane

A multi-tenant platform for running AI agents in isolated [Vercel Sandboxes](https://vercel.com/docs/sandbox), exposed via a REST API. Supports any model via [Vercel AI Gateway](https://vercel.com/docs/ai-gateway) with dual runner architecture ([Claude Agent SDK](https://docs.anthropic.com/en/docs/claude-code/sdk) + [Vercel AI SDK](https://sdk.vercel.ai)).

**AI Agents as an API** — one POST request to launch an agent with skills, MCP connectors, streaming, and full observability. Built for platforms.

Features

• Any model, any provider — supports all models via Vercel AI Gateway; dual runner: Claude Agent SDK for Anthropic models, Vercel AI SDK for everything else
• Isolated sandboxes — every execution spins up a Vercel Sandbox from a pre-built SDK snapshot (~3s cold start) with its own filesystem, network policy, and resource limits
• Skills & plugins — inject custom skills and tenant-scoped marketplace plugins into agents before execution
• Scheduled executions — configure agents to run on a schedule (hourly, daily, weekdays, weekly) with timezone-aware execution
• Multi-tenant — row-level security, per-tenant API keys, budget controls, and rate limiting
• MCP connectors — connect agents to external tools via Composio (GitHub, Slack, Firecrawl) or tenant-scoped custom OAuth 2.1 MCP servers
• A2A protocol — expose agents to external A2A-compliant clients (LangGraph, CrewAI, Semantic Kernel, etc.) via Agent Cards and JSON-RPC; streaming SSE, idempotency, budget clamping
• Full observability — every session message stores a transcript, token usage, cost, duration, and trigger source (API/schedule/playground/chat/A2A/webhook)
• Playground — test agents interactively from the admin dashboard with real-time event streaming
• SoulSpec v0.5 identity — full SoulSpec compliance: SOUL.md (Personality/Tone/Principles), IDENTITY.md (Name/Role/Creature/Emoji/Vibe), STYLE.md, AGENTS.md, HEAR