Platos

Build, ship, and operate AI agents on infrastructure you own. Apache 2.0. Self-hostable in one docker compose up.

What is Platos?

Platos is a complete agent runtime — the open-source replacement for hosted services like Claude Managed Agents and OpenAI Assistants.

It bundles everything an agent needs to live in production:

• Streaming chat runtime with prompt caching, tool-calling, structured outputs, sub-agents, and multi-turn compaction.
• Durable execution layer built on top of trigger.dev — every long-running tool call, scheduled job, or batch operation is a resumable run with retries, queues, and traces.
• Universal MCP gateway that federates four tool families (entity-pushed, native, skills, control plane) behind a single endpoint with OAuth scoping and per-tool ACL.
• Memory + skills + observability wired together at the runtime layer — vector store, knowledge graph, manifest-driven skills, OpenTelemetry traces, ClickHouse cost ledger.
• Multi-tenant scope model — every row is keyed by (organizationId, projectId, environmentId). Build SaaS or run an internal platform; the same primitives fit both.
• BYOK for every provider — Anthropic, OpenAI, Google, Vertex AI, OpenRouter. Keys are encrypted in your database and never leave it.

You own the infra, the data, and the model choice. No vendor seat license, no telemetry-based billing, no lock-in.