xAI shipped v0.2.97 of Grok Build, its terminal coding agent, on July 11, 2026. If you have not tried Grok Build yet, it is xAI's answer to Claude Code and Codex CLI: an agentic CLI that plans, edits, runs shell commands, and iterates against test output inside your actual working tree, rather than just answering questions about code in a chat window. This release is smaller than a headline feature drop, but it fixes several things that mattered if you were running Grok Build in scripts, on SSH, or on a devbox that kept freezing.
The features that actually change your workflow
The biggest addition is cost visibility. Headless JSON output now includes token usage and cost per prompt and per session, and the SDK exposes the same data through Turn.usage. If you have been running Grok Build inside CI or a batch script and guessing at spend from your xAI billing dashboard after the fact, you can now pull exact numbers programmatically, per call. That matters more than it sounds: teams running agentic coders at scale need per-task cost attribution to decide which workflows are worth automating and which are burning tokens on trivial edits.
Two changes affect anyone working interactively:
- Edit and resubmit from the transcript. Double-clicking or pressing Enter on a previous user message now lets you edit it and resubmit directly, instead of forcing you to scroll up, copy the text, and retype a corrected prompt.
- Ripgrep no longer prompts for permission. Shell commands using
rgrun without a confirmation dialog by default. This is a small change but a real one if you lean on ripgrep for code search inside agent sessions, since permission prompts were previously a friction point that broke flow on every single search.
Voice mode is now available for API-key sessions too, not just OAuth-authenticated accounts. And there are new environment variables for tuning scroll and draw cadence, aimed at people running Grok Build on high-refresh-rate displays where the default redraw rate looked choppy or lagged input.
Bug fixes worth knowing about
The fix list reads like a list of things that were genuinely annoying in daily use:
| Fix | Why it mattered |
|---|---|
| Background tasks in headless mode now killed on exit | Previously leaked processes after a headless run finished |
| Agent process leaks on failed spawns fixed | Failed launches and missing-stdio teardown no longer left orphaned processes |
| MCP and skills reload freezes fixed | Frequent reloads could hang whole sessions on devboxes |
| HTTP-based MCP servers now auto-recover | Tools like the Slack-focused nebula server survive disconnects instead of dying silently |
| iTerm2-over-SSH clipboard copy fixed | Copy previously reported success even when it silently failed |
The MCP auto-recovery fix is the one I'd flag if you run any HTTP-based MCP server in a long session. A dropped connection used to mean restarting the whole agent session to get your tools back. That's now handled automatically.
How this fits into the bigger picture
Grok Build is still in beta and gated behind SuperGrok Heavy, SuperGrok, or X Premium+ subscriptions, with pricing reported around $300/month at list (or a $99/month intro rate), which puts it in the same bracket as a heavy Claude Code or Codex CLI seat rather than a pay-as-you-go API tool. The underlying model, Grok 4.5, has posted competitive numbers on agentic coding benchmarks: on SWE Bench Pro, Grok 4.5 reportedly scored 64.7% versus Opus 4.8's 69.2%, and on Terminal Bench 2.1 the gap between top models is now within a percentage point or two rather than the wide spreads seen a year ago.
What v0.2.97 tells you is where xAI is focused right now: not new agent capabilities, but operational maturity — cost accounting, connection resilience, and cleanup of leaked processes. That is exactly the kind of unglamorous work a CLI needs before teams will trust it in CI/CD pipelines rather than just local exploratory sessions.
Practical takeaway
If you already run Grok Build, update now for the MCP auto-recovery fix alone if you use any HTTP MCP server. If you are evaluating it against Claude Code or Codex CLI, this release is a signal that xAI is treating Grok Build as production tooling rather than a demo, but the subscription-gated access and beta status still mean it belongs in exploratory or local workflows before you put it on your critical path.
