Tribe AI

Tribe AI is an open-source, MIT-licensed low-code tool for building and coordinating multi-agent teams, built on the LangGraph framework. Users design teams visually with drag and drop, then chat with them to solve tasks. The README notes the project is under heavy development and significant changes may occur.

What It Is

Tribe lets you split a complex task across agents that each specialize in a different area. Teams are created on a Teams page, where members are added by dragging nodes, and are then tested from a Chat tab. Example use cases in the README include football analysis, trip planning and customer service teams.

Workflow Types and Skills

Tribe supports two workflow styles. Sequential workflows run agents one after another in a deterministic order. Hierarchical workflows use a team leader that breaks tasks down, delegates them to members or sub-team leaders, and decides whether to return a response or delegate further.

Agents can be equipped with skills. Built-in skills are duckduckgo-search, wikipedia and yahoo-finance. Custom skills can be defined as JSON function definitions for simple HTTP requests, or written as LangChain tools in the codebase.

Knowledge, Models and Oversight

Retrieval Augmented Generation lets users upload private data and choose which data goes into an agent's knowledge base. The default embedding model is BAAI/bge-small-en-v1.5 and it can be changed through an environment variable. Open-source models can be used through Ollama or any endpoint compatible with the OpenAI chat completions API. In sequential workflows, a skill can require human approval before it runs, and the reviewer can reject the action with an optional message.

Deployment

Tribe is deployed with Docker, either locally or on a remote server. Secrets such as SECRET_KEY, FIRST_SUPERUSER_PASSWORD and POSTGRES_PASSWORD must be changed before deployment. Public API endpoints allow external applications to interact with teams, and LangSmith is supported for observability.