Skip to content

Evokai

Evokai, a small mage in a big violet hat

Summon AI agents into any app, with the governance a company needs. Any model, from the big clouds to your own servers; your app's tools over MCP, skills and voice; always acting as the signed-in user, within budgets, with an audit trail.

Every product ends up wanting an assistant that can do things: read the data on screen, run the app's own actions, answer by voice, and never step outside what the user is allowed to do. Evokai is that layer, written once. An application connects with a key, describes its agents in a YAML file (or in the admin console), and its users talk to them through a small headless client. The application keeps its identity, its data and its interface; Evokai brings the agents.

Experimental. Evokai is early-stage. Interfaces (the HTTP API, the configuration, the clients) may change between 0.x releases, and a release may need a manual migration step; read the changelog before you upgrade.

Start here

Quick start Evokai, its console and its own Postgres with Docker Compose, in five minutes
Features Everything Evokai does, in one table
Architecture Concepts, the protocol, voice, limits, storage and the hard rules
Configuration Environment variables, application YAML, bootstrap, keys, time to live
Deployment Running it, Docker, database setup, CI gates, images, SBOMs and signing
Security Threat model and the security layer's controls, mapped to the OWASP LLM Top 10
Data protection Residency, jurisdiction and retention rules for every processor
Connectors Ready MCP servers: SQL, REST/OpenAPI, Google Drive, SharePoint and OneDrive, IMAP

In your app

Your backend mints a short session token for the signed-in user with its server key, and the page talks to the agent with it, through @evokai/evokai-client (headless) or the ready-made panel and voice stage of @evokai/evokai-ui. Python backends use the evokai-client package. A complete app is in examples/bookshop-app.

The source, the issues and the releases are on GitHub. Evokai is licensed under the Apache License 2.0; the name, its logo and its mascot are not covered by the license (see TRADEMARKS.md).