Anima
A self-hosted browser runtime that lets an AI agent work in a signed-in browser.
Anima is a self-hosted browser runtime for AI agents. Through an MCP server it gives an assistant a real, signed-in browser to work in, so a task starts on the page where the work is, past the login screen.
What it is
A self-hosted browser runtime for AI agents, built solo. Each profile is a full Firefox with its own sessions, and each identity signs in with its own password vault account. An AI assistant drives a profile through an MCP server with real clicks and keystrokes, so a task starts in a signed-in browser instead of at a login screen, and a profile can be bound to a project so each agent works in its own browser.
Built with
What you can check
- An MCP server exposes the browser to an assistant as tools
- Credentials are typed by the runtime and never pass through the model
- One profile per project, so two agents never share a session
The problem: an agent stops at the login screen
Most useful work on the web sits behind a sign in. An AI agent with a fresh, empty browser spends its effort on login forms and second factors, and the usual shortcut is handing it a password. Anima gives the agent a browser that is already signed in and keeps the password away from the model.
Each profile is a full Firefox with its own sessions, kept between runs. What an agent signed into yesterday is still signed in today, the way it is for a person who never closes their browser.
Credentials never pass through the model
When a sign in is needed, the runtime types the credential itself, from a password vault, with real keystrokes. The model asks for the login to happen and never sees the secret, so the password is absent from the prompt, from the tool arguments and from the transcript.
Each identity signs in with its own vault account, so an agent can reach only the credentials that belong to the identity it is running as.
A secret that reaches a model's context can be repeated in its output, kept in a provider's logs or carried into a later task. Keeping it out is a property of the design, and it holds whether or not the model behaves.
An MCP server, and one browser per project
The browser is exposed to an assistant as tools through an MCP server: open a page, read it, click, type, upload a file. A profile can be bound to a project, so each agent works in its own browser and two agents never share a session.
The tools act on the page with real clicks and keystrokes, the same input a person gives. After each action the assistant gets back what changed on the page, which keeps a long task inside its context.
Anima is private and in daily use on my own projects. The pattern is what I bring to client work: an MCP server with a stated boundary for every tool, and secrets that never enter a model's context.
Questions
- What is a browser runtime for AI agents?
- Software that runs and keeps real browser profiles, and lets an AI assistant drive one through tools. The assistant opens pages, reads them, clicks and types, in a browser that keeps its sessions between tasks.
- Does the model ever see a password?
- No. The runtime reads the credential from a password vault and types it into the page itself. The model only asks for the sign in to happen, so the secret is never in a prompt, a tool argument or a transcript.
- Is Anima available to use?
- No. It is a private project, built solo and used on my own work. The same design, an MCP server with a written boundary for each tool, is what the MCP server service on this site delivers for a client's product.
Facts on this page verified 2026-11-10