AI Agents / ZooGent
Open source

Build agent teams
with ZooGent

ZooGent is our open-source toolkit for building AI agents and assembling them into a working team. Design the team once, then run it on CrewAI, LangGraph, the OpenAI Agents SDK, or ZooGent's own runtime.

ZooGent TOOLKIT Research Writing Analysis Review Scout Triage Sync QA
ZooGent AGENT-TEAM CREATOR builds THE AGENT TEAM A Agent 1 Claude B Agent 2 GPT-4 C Agent 3 Gemini runs on Orchestrator CrewAI · LangGraph · OpenAI Agents SDK

What ZooGent is

ZooGent is not another AI. It is the toolkit you build the team with. Each agent gets a role, the skills to do it, memory of what worked, and a way to hand work to the next agent. You describe the team; ZooGent turns it into running code.

Agents, not one big prompt. Each agent has one job and the model that fits it. Small, testable, swappable.
The design is the asset. Roles, skills, and flows live as plain files. That is what you own, and what moves between frameworks.
Built to run cheap. Prompt caching, a cheap model by default with an escape to a stronger one, selective context. A naive setup costs several times more.

What every agent gets

A single chatbot forgets and works alone. ZooGent gives every agent four things, and that is what turns a pile of prompts into a team.

01

Tasks

One agent hands work to another with clear instructions. Like a direct message between coworkers.

Agent to Agent
02

Team Knowledge

A shared board where agents post what they learn. Once approved, the whole team can see it and use it.

Collective Intelligence
03

Memory

Each agent remembers what worked and what did not. A scoring system keeps useful memories close and lets the rest fade.

Personal Experience
04

Skills

Instructions written by people and shipped with the agent. Read-only. This is what makes a cheap model do the job of an expensive one.

Human Instructions

One team design. Any runtime.

Frameworks come and go. ZooGent keeps your team design — roles, skills, flows — separate from the framework underneath. It runs on its own runtime today, and is built to be orchestrator-agnostic.

ZooGent Runtime

Runs today

Self-hosted and cost-tuned. Skills, memory, team knowledge, tasks, and cost tracking built in. This is what every Devvela agent team ships on.

CrewAI

Role-based crews

One of the most-used agent frameworks, built around roles and tasks. A target runtime, so a team you design in ZooGent is not tied to ours.

LangGraph

Stateful graphs

Graph-based orchestration with checkpointing. A target runtime for teams that need fine-grained control over state.

OpenAI Agents SDK

Sessions & handoffs

OpenAI's framework for agents, sessions, and handoffs. A target runtime, so your team design follows the ecosystem.

Open, self-hosted, yours

ZooGent is open source under the MIT license. You can read every line, self-host it, and use it for commercial work. We maintain it, and we build every Devvela agent team on it.

MIT license Self-hosted TypeScript SQLite storage Docker-ready Any LLM provider
View on GitHub

FAQ

ZooGent is Devvela's open-source toolkit for building AI agents and assembling them into a working team. Each agent gets a role, the skills to do it, memory of what worked, and a way to pass work to the next agent. You describe the team; ZooGent turns it into running code.

Yes. ZooGent is open source under the MIT license. You can self-host it, read every line, and use it for commercial work. Devvela maintains it and builds client agent teams on it.

ZooGent keeps your team design separate from the framework that runs it. It runs on its own self-hosted runtime today, and is built to be orchestrator-agnostic, with CrewAI, LangGraph, and the OpenAI Agents SDK as the runtimes it targets, so a team is not locked to one framework.

ZooGent is a developer toolkit and agents are written in TypeScript. If your team does not build it in-house, Devvela designs and ships the agent team for you, usually with a first working version in about two weeks.

ZooGent is built for cost-efficiency: prompt caching, a cheap model by default with an escape to a stronger one only when needed, and selective context loading. Together these cut the running cost of an agent team several times over compared with a naive setup.

Want a team built on ZooGent?

Tell us the workflow you want to automate. We design the agents, build the team, and run it. Most setups reach a first working version within two weeks.

Book a 15-min call