agent-browser has added a new “chat” mode that reframes the tool less as a scriptable browser controller and more as a browser agent—with both one-shot CLI usage and an interactive loop. The update centers on a single entry point, agent-browser chat, plus a built-in AI chat surface in the dashboard that can run agent-browser commands directly.
From commands to conversations (without losing the CLI)
The headline change is the new chat interface, which supports two patterns:
- One-shot runs, such as:
agent-browser chat "open google, search for dogs" - Interactive sessions that start with:
agent-browser chat
In addition, the dashboard now includes built-in AI chat, positioned as a more conversational way to drive the same underlying command set. Chris Tate also notes that chat can execute any agent-browser command, and that the tool can be used as a sub-agent—a clue that this is meant to plug into broader agentic workflows rather than remain a standalone automation utility.
Install and upgrade
The install path is intentionally simple:
- First-time install:
npm i -g agent-browser - Upgrade:
agent-browser upgrade
Auth, profiles, and keeping sessions around
A common sticking point for browser agents is authentication. In replies, Tate outlined multiple options for handling credentials and state in agent-browser:
Reusing Chrome and persisting sessions
- Reuse existing Chrome login sessions via
--profile Default - Auto-save/restore cookies + storage across runs using
--session-name github
Storing credentials and moving state between runs
- An encrypted credential vault with commands like
auth save mysiteandauth login mysite - Portable auth snapshots using
state saveandstate load
Working with headers and basic tokens
- Tokens and basic auth via
set headersandset credentials
Taken together, these cover the practical cases teams hit quickly: staying logged in, not redoing MFA flows every run, and carrying state between environments.
Developer questions are already shifting to the next layer
The thread quickly moved past “can it click things” into more operational concerns: token usage efficiency, exposing the dashboard over the network (0.0.0.0), whether it can access console/network tabs, and how it stacks up against Chrome’s MCP in terms of flexibility and control. There were also early asks about E2E testing workflows (Playwright-style) and even “templates support” for repeatable higher-level tasks.
For now, the concrete shipped change is clear: agent-browser chat turns the CLI into a conversational interface, while keeping direct command execution and session tooling close at hand.