Factory AI

Factory AI is a multi-model AI platform for developers that standardizes usage and billing across different language models. Instead of buying capacity directly from each provider, you consume Factory Standard Tokens, which are mapped to specific models via multipliers and an internal caching system. The service targets engineering teams and power users running substantial LLM workloads, especially for coding and automation.

A key component is Factory Droid, an AI agent that integrates into the Zed code editor. You install the Factory Droid Zed extension, authenticate via a device-code login flow, and then start agent sessions from Zed’s Agent Panel. Beyond the editor integration, Factory offers a web app (for signup, billing, and API keys), a CLI, and an HTTP API (with endpoints like machine template management) so you can integrate the platform into your own tools and workflows.

Factory’s main strength is giving you predictable, model-agnostic pricing via Standard Tokens and per-model multipliers. According to their documentation, cached tokens are billed at one-tenth the rate of standard tokens, so typical workloads that benefit from caching can stretch their quota further than direct on-demand calls to model providers. You also get access to a curated set of models (including internal Droid Core and external models like GPT, Claude, and Gemini) under one account, which simplifies experimentation and switching.

On the agent side, Factory Droid runs as an agent in Zed, with support for starting and resuming sessions from within the editor. The docs reference "Models and Autonomy" and optional MCP (Model Context Protocol) servers for additional context, indicating support for more agentic behavior, but the exact level of autonomous action and available tools is not fully detailed in the snippets provided. Setup assumes developer familiarity: you’ll be working with API keys, environment variables, and a CLI (which currently does not support Windows on ARM). There’s no explicit information in the provided docs about data privacy, persistence of code or chat history, or enterprise governance features, so if those are critical you’ll need to review their full documentation or contact them. Non-technical users or teams looking for a turnkey, non-configurable assistant are likely not the target audience.