vercel-labs/agent-browser v0.41.0 on GitHub

@agent-browser/eve tools now come from the MCP tool catalog. The extension exposes the CLI's MCP tools instead of 21 hand-written ones, so tool names and inputs change: navigate is open, evaluate is eval, press_key is press, select_option is select, set_checked is check and uncheck, get is get_text, get_url and the other get_* tools, tabs is the tab_* tools, and wait_for is the wait_* tools. The same names apply to @agent-browser/eve/tools overrides. The sandbox needs agent-browser 0.41, which the extension installs by default (#2241)

Added agent-browser mcp call <tool> '<json>' to run one MCP tool without a stdio session and print its result as JSON with the agent-browser version, for runtimes that cannot keep a process open between tool calls (#2238)
Added @agent-browser/sandbox/mcp-tools.json, every MCP tool with its title, description, input schema, annotations and profiles, generated from the CLI and checked by a test (#2240)

@agent-browser/eve runs each tool as its own short-lived process, so no agent-browser process outlives a tool call. The 28 core tools are listed to the model and the other 14 are deferred, found with eve__search, which keeps the listed tool context at its previous size (#2241)

MCP tool arguments can no longer set global flags. The CLI reads global flags anywhere in its arguments, so a value such as a selector or extraArgs entry could carry its own --allowed-domains and override the server or caller allowlist. Tool calls that try are now rejected (#2238)