Foundry

Why we included this project

Most agent frameworks leave the production layer to you. Foundry bundles it: define the agent with a single call, put its tools on an MCP server, and flip on memory, security scanning, response caching, and tracing through one parameter each. The runtime keeps the agent running as a recoverable process that wakes on a schedule or in response to a webhook or a file change, with caps on tool calls and spending and a human approval gate for risky steps. Tenant isolation is structural rather than a flag you remember to set, so memory, cache, and audit logs stay separate per customer by construction. If you would rather audit one stack than glue together several libraries, and it has to talk to OpenAI, Anthropic, Gemini, or a local model, this is worth a look.