Overview

Agents Gateway is a FastAPI-based platform for managing, orchestrating, and executing LLM-powered agents. It provides a unified REST API for agent CRUD, multi-agent teams, knowledge management, token storage, prompt templates, skills, and a supervisor/worker execution platform.

Architecture

                                    +-------------------+
                                    |   Client / UI     |
                                    +--------+----------+
                                             |
                                   X-API-Key | (V2 routes)
                                             |
                              +--------------v--------------+
                              |      agents-gateway         |
                              |    FastAPI  (port 8000)     |
                              |                             |
                              |  /v2/agents   /v2/teams     |
                              |  /v2/prompts  /v2/knowledge |
                              |  /v2/tokens   /v2/skills    |
                              |  /v2/engines  /v2/targets   |
                              |  /v2/approvals              |
                              |  /admin       /health       |
                              +-+----------+----------+-----+
                                |          |          |
                     +----------v--+  +----v----+  +--v--------+
                     | PostgreSQL  |  | Qdrant  |  |  Prompts  |
                     | (metadata,  |  | (vector |  |  Storage  |
                     |  sessions,  |  |  KB)    |  | (postgres)|
                     |  tokens)    |  +---------+  +-----------+
                     +-------------+

Key Concepts

• Agent: An LLM-powered assistant defined by a prompt template, model, and configuration. Stored in PostgreSQL with templates in the prompts storage backend.
• Team: A group of agents that collaborate. Supports coordinate mode (round-robin collaboration) and supervisor mode (leader classifies, workers execute).
• Prompt: A reusable prompt template stored in the prompts storage backend and referenced by agents.
• Knowledge: Document-based knowledge bases with dual-level isolation (tenant and collection) backed by Qdrant vector search.
• Token: Encrypted OAuth2/API key credentials stored per user per integration (Google, Slack, OpenAI, etc.).
• Skill: A reusable instruction set with references and scripts that agents can leverage.
• Engine: An execution engine registry entry (e.g., code_agent via Claude Code, managed_agent via API, direct_ops).
• Target: An execution target (local machine, SSH host, remote service, managed agent API).
• Approval: A human-in-the-loop approval request for tool calls that require confirmation.
• Pack: A YAML-defined bundle of agents, prompts, and team configuration that can be applied as a unit.