Developers hitting Anthropic's usage caps now have a new shield: cc-limit-pacer, a Python-based tool designed to automate pacing against Claude Code's 5-hour and weekly plan limits. Released by Rahul Balakavi and hosted on GitHub, the utility addresses the frequent frustration of being locked out mid-session by intelligently managing model selection and batch execution. It is currently at version 0.1.1 and requires Python 3.9+ with standard library dependencies only.

Automated Model Tiering and Batch Holds

The core mechanism relies on hooks that monitor usage rates in real time. When a session is deemed "hot"β€”defined as a usage window being at least 50% full and on pace to hit its limitβ€”the tool automatically steps down the model tier for new sessions. This means switching from Fable to Opus, or Opus to Sonnet, while simultaneously disabling the advisor model to reduce token consumption. Crucially, this pacing never interrupts interactive sessions; instead, it holds automated runs, such as those initiated via the SDK or claude -p, until the usage window resets.

Extended Context Windows and Audit Trails

Beyond simple pacing, cc-limit-pacer optimizes context management by adjusting auto-compact thresholds. On 1M-window models, sessions are configured to auto-compact at approximately 830k tokens rather than the default 570k. This allows developers to retain more context and reduce the frequency of compaction events, which are often a source of data loss and performance lag. The tool also includes an audit feature that calculates potential savings, tracking wasted resources like unused weekly allowances, Fable capacity, and advisor costs, presenting this data as text or a dedicated stats page.

Installation and Calibration Requirements

Installation is handled via the Claude Code plugin marketplace, though the underlying repository is currently private, requiring users to have specific GitHub access or authenticated git credentials. Users must run a calibration sequence using /cc-limit-pacer:calibrate with their current 5-hour and weekly usage percentages and the weekly reset time, all derived from the /usage command. The tool also offers a simulate command to replay past usage history under different policies, providing evidence of how the pacer would have altered lockout frequencies and hours lost.

Key Takeaways

  • The tool prevents lockouts by downgrading model tiers and holding batch jobs when usage exceeds 50% of the limit.
  • Auto-compaction is delayed to ~830k tokens on 1M-context models to preserve more conversation history.
  • Calibration requires manual input of current usage stats and reset times from the Claude Code interface.
  • The repository is private, requiring explicit GitHub access for installation via the plugin marketplace.

The Bottom Line

cc-limit-pacer is a necessary stopgap for developers who treat Claude Code as a production dependency rather than a casual toy. By automating the trade-off between model quality and availability, it transforms unpredictable lockouts into managed, lower-quality sessions, which is a far superior outcome for sustained development workflows.