promptise-com/Foundry

The foundation layer for agentic intelligence.

What it solves

Promptise Foundry is a full-stack agentic engineering framework designed to eliminate the need for developers to manually assemble a fragmented stack of separate libraries for building AI agents. It provides a unified system for creating agents, the tools they use, their reasoning logic, and the production-ready infrastructure required to deploy them to customers securely.

How it works

Foundry integrates several core components into a single framework:

  • Agent Core: A single function (build_agent()) that connects models to tools, memory, security scanners (to block prompt injection and redact PII), and response caching.
  • Reasoning Engine: Allows developers to define an agent's thinking process as a readable graph with presets for common patterns like research, debate, and plan-act-reflect.
  • MCP Server SDK: A tool-building kit that turns Python functions into Model Context Protocol (MCP) tools with built-in authentication, rate limits, and audit logs.
  • Agent Runtime: A system for managing long-running agents that can be triggered by schedules, webhooks, or file changes, featuring crash recovery and budget limits.
  • Prompt Engineering: A system for treating prompts as versioned, typed blocks of software rather than simple strings.
  • Agent Identity: Provides agents with their own authenticated identities (via OIDC, AWS, GCP, etc.) to replace shared API keys.

Who it’s for

It is built for developers and teams who want to move from an AI agent prototype to a production-grade, multi-tenant system that is secure, observable, and governed without rebuilding the production layer from scratch.

Highlights

  • Native Multi-tenancy: Structural data separation for memory, cache, and conversations per customer.
  • Enforced Human Approval: Server-side gates that require human sign-off for risky tool calls.
  • MCP Native: Tools are compatible with Promptise agents as well as other MCP clients like Claude Desktop and Cursor.
  • Air-gapped Capability: Supports local security models, embeddings, and vector stores for offline or regulated environments.
  • Recoverable Execution: A journal system that allows agents to resume from where they stopped after a crash.