Unified Tool-Calling & MCP Platform for AI Agents

Overview

A tool-calling and Model-Context-Protocol (MCP) platform that exposes a large catalog of third-party SaaS tools to AI agents through a single, authenticated, intent-aware interface. The codebase is an extensively extended fork of the open-source ACI.dev platform, adapted and operated as the integration backbone for an internal agentic product.

Why It Exists

Wiring AI agents to real-world tools means writing bespoke OAuth flows, API clients, and permission logic for every service. This platform centralizes that work: agents get unified, secure function calls (or a single MCP endpoint) instead of dozens of one-off integrations, with multi-tenant auth and dynamic tool discovery handled centrally.

What We Built

Building on the open-source ACI.dev base, the team contributed dozens of custom commits across the backend and frontend. The FastAPI backend hosts a deep integration library, 100+ tool “apps” under backend/apps/ (Gmail, Slack, Notion, Jira, GitHub, Stripe, Supabase, Vercel, Google Workspace, Firecrawl, Tavily, crypto/data APIs, browser-automation providers, and many more), an Alembic-migrated PostgreSQL schema with pgvector for semantic tool search, PropelAuth multi-tenant authentication, Stripe billing, and Sentry/Logfire observability. A Next.js frontend provides the management console; Docker Compose orchestrates server, runner, and database services.

Technologies & Approach

Python/FastAPI with SQLAlchemy + Alembic and pgvector for embedding-based tool discovery; OpenAI and LangChain for the agent/tool-calling layer; an MCP server interface so any MCP-compatible IDE or agent can consume the catalog. PropelAuth, Stripe, Sentry, and Logfire round out the production concerns.

Outcome / Impact

Demonstrates the studio’s ability to take a fast-moving open-source agent-infrastructure project, extend it with substantial custom integration and auth work, and run it as the connective tissue for a real agentic product. Proves capability in MCP, multi-tenant tool-calling architecture, and large-scale third-party API integration.

Capabilities Demonstrated

• Building and operating agentic tool-calling infrastructure (MCP + direct function calls)
• Extending a complex open-source platform with production auth, billing, and observability
• Large-scale third-party API / SaaS integration (100+ connectors)
• Semantic tool discovery with vector search over PostgreSQL/pgvector
• Multi-tenant authentication and granular permission models for AI agents