# Veldra

A self-hostable, local-first platform that compiles natural-language descriptions into working AI agents, complete with tools, MCP connectors, a RAG knowledge base, and multi-agent teams, then lets users edit them by talking to them.

Veldra is a self-hostable, local-first agent-harness platform. Describing a need in plain language compiles it into a working agent: its policy, tools, skills, a RAG knowledge base, a reasoning method, or even a whole team of agents. Every agent is stored as a versioned specification record in a database, so building an agent means compiling a request into a validated spec, and changing it means approving a patch; the runtime only interprets that spec.

A composer routes a typed task to the best existing agent or builds a new one on the fly. Built-in tools such as a calculator, web scraper, RAG search, and HTTP client are always available, and external MCP connectors can be installed from a template gallery, with side-effecting tools requiring explicit approval. Knowledge bases support uploaded documents or indexed web pages, several retrieval modes, and a choice of vector store, and a visual workflow builder wires together nodes such as classifiers, tool calls, and conditionals.

Agents can be organized into teams with a coordinator and specialists, and each has a persona that shapes its tone. With auto-improve enabled, an agent stores lessons from negative feedback or its own mistakes and adjusts accordingly.

Website: https://github.com/shamspias/customizable-gpt-chatbot