Skip to content
News Product launch

Claude Code 2.1.198 makes Claude in Chrome generally available and adds auto-commit background agents

· by Pondero Newsdesk

The short version

Anthropic shipped Claude Code 2.1.198 on July 1, 2026, graduating Claude in Chrome from preview to general availability and giving background agents the ability to commit, push, and open a draft pull request without stopping for confirmation.

Claude Code 2.1.198 makes Claude in Chrome generally available and adds auto-commit background agents

Background agents in Claude Code can now commit their own work and open a draft pull request when they finish, removing the confirmation step that previously made them pause mid-task. That change arrived in version 2.1.198, released July 1, alongside Claude in Chrome reaching general availability, per the official Claude Code changelog.

What shipped

Two changes in 2.1.198 stand out for teams running Claude Code at any scale.

Claude in Chrome moved from preview to general availability. Any Chrome user can now run Claude Code agents with full browser access without needing a terminal. This removes the biggest distribution hurdle for the feature: previously, using Claude in Chrome required a developer-configured terminal environment, which kept it out of reach for most non-engineering roles in a team. GA status means Anthropic treats it as production-ready and supports it with the same stability expectations as the core CLI.

The auto-commit behavior for background agents is the second major change. Agents launched from the claude agents agent launcher now commit their changes, push to the remote branch, and open a draft PR automatically when they finish work in a worktree. Previously they stopped and waited for user confirmation before taking those steps. Per the changelog, that behavior was opt-in in earlier builds and is now the default path.

The release also shipped notification hooks for background agents: agent_needs_input and agent_completed now fire in the Notification hook system when a session pauses for input or completes, giving teams a programmatic way to track agent status without polling. A /dataviz skill for chart and dashboard design guidance with a runnable color-palette validator also landed in the same build.

On the infrastructure side, Anthropic added Claude Platform on AWS (anthropicAws) as an upstream provider in Gateway, and model-not-found responses now advance the failover chain rather than returning an error to the caller. Version 2.1.199, released July 2, layered on stacked slash-skill invocations: '/skill-a /skill-b do XYZ' now loads all leading skills (up to five), not just the first.

Why it matters

The auto-commit change matters most for teams that have been using background agents for exploratory tasks. Until now, the workflow was: launch agent, let it write code, then manually review, commit, and push before the branch existed for other reviewers to see. That last step required someone to be at the keyboard. With auto-commit on by default, an agent launched in a worktree can complete a full branch and post a reviewable draft PR without any human step between task assignment and code review.

Teams with review-before-push policies should take note. The commit lands on the remote before a human sees the diff. Draft PR status prevents auto-merge, but the code enters the branch history regardless. Teams using background agents in repositories with strict push controls should confirm their branch protection rules account for agent pushes.

Claude in Chrome GA opens a separate question about adoption. Agent tasks that benefit from live browser interaction, like UI testing, research synthesis, or form-filling workflows, become available to anyone with Chrome, not just engineers comfortable with terminal tooling. Anthropic has not published usage figures from the preview period.

Context

The 2.1.198 release included a cluster of reliability fixes for the background-agent runtime: stuck "Running" status after sessions complete or resume, agent teammates that die on API errors now reporting failure instead of going silent, and automatic retry with backoff for transient ECONNRESET drops. The Explore agent also now inherits the main session's model rather than running on Haiku by default, capped at Opus.

What to watch next

Two things are worth tracking. First, whether Anthropic publishes adoption data for Claude in Chrome now that it has reached GA, which would reveal whether no-terminal browser access actually pulls in non-engineering use cases or whether adoption stays concentrated among developers. Second, how teams with existing CI/CD pipelines respond to auto-commit agents: the pattern requires either trust in the agent's commit behavior or new branch protection rules that gate on automated pushes explicitly.

Sources