Vibemetric – An open-source app that grades your Claude Code sessions

Vibemetric launched an open-source macOS application designed to inspect and grade local Claude Code and Codex terminal sessions directly on a developer workstation. According to vibemetric.app's report, the utility evaluates how autonomous coding agents interact with repositories, uncovering patterns where agents repeat basic configuration mistakes, skip automated test suites, or burn through high-tier model budgets for trivial file edits. The application operates entirely offline on the local machine without external telemetry, parsing session transcripts already written to disk and scoring overall developer-agent workflows out of one hundred points across ten operational disciplines.

This analysis breaks down the architecture, scoring metrics, and remediation templates shipped in the initial release of Vibemetric for software engineers managing agentic coding workflows. It examines how engineering teams running Claude Code and Codex can diagnose unseen token waste, prevent broken code from silently entering production branches, and automate governance rules through structured configuration files. The breakdown details concrete session findings, empirical benchmarks on agent performance, step-by-step model routing configurations, and installation commands necessary to establish systematic oversight over local terminal agents.

What shipped in Vibemetric – An open-source app that grades

The initial release of Vibemetric delivers a native macOS client that scans local terminal logs from Claude Code and Codex without uploading session data to external servers. The desktop interface presents an aggregate score from zero to one hundred alongside granular ratings across ten foundational categories: Planning and Specs, Context Engineering, Memory and Knowledge, Testing and Evals, Efficiency and Cost, Tool Integration, Task Mapping, Feedback Loops, and Security. Rather than requiring developers to manually fill in performance logs, the software automatically discovers existing session transcripts, aggregates past command execution histories, and benchmarks real agent behavior over daily rolling windows.

Each scored category pairs historical session transcripts directly with concrete prompt recommendations that engineers can insert into project configuration files. When the scanner uncovers repetitive failures, such as agents executing changes without triggering verification tests or running tasks under unrestricted system permissions, it provides single-click prompt templates tailored for repository instructions. The tool groups project metrics by repository workspace, allowing developers to monitor improvement trajectories over time, compare multi-session runs, and implement standardized definition of done checklists that enforce automated test execution before any task is marked complete.

What improved in Vibemetric – An open-source app that grades

The diagnostic engine surfaces measurable operational gaps documented across modern software engineering research, providing calibrated baselines against widespread agent inefficiencies. Industry studies highlighted by the platform reveal that unconstrained agents generate wide operational swings, including up to thirty times token cost variance across identical tasks according to the Stanford Digital Economy Lab. Analysis across six hundred twenty-three million code changes by GitClear similarly documented an eighty-one percent spike in duplicated code blocks alongside a seventy percent collapse in code refactoring since 2022. By auditing session transcripts against these failure patterns, Vibemetric quantifies how often agents bypass testing or run redundant edits.

Applying the diagnostic fixes produces measurable performance gains across multi-file engineering tasks, lifting baseline audit scores from forty-six to eighty-three points. In audited developer workspaces, unguided agents frequently skipped test runs after twenty-three out of thirty-one code modifications, leading to silent production regressions that required manual rollbacks days later. Implementing structured project rules eliminates repeated instruction prompts, which previously recurred across nine separate sessions for basic package manager commands. Furthermore, enforcing explicit model routing redirects routine typo corrections and variable renames away from high-end frontier models, reclaiming up to one-third of total weekly token expenditure.

What you gain from Vibemetric – An open-source app that grades

Developers deploying the diagnostic tool gain full transparency into hidden operational costs and execution vulnerabilities inside local terminal sessions. The diagnostic breakdown exposes scenarios where agents execute commands with all permission checks disabled, preventing malicious or accidental modifications to files outside the designated project root. By auditing security discipline alongside task completion rates, teams address critical vulnerabilities highlighted in the Veracode GenAI Code Security Report, where autonomous code samples showed an average security pass rate of only fifty-six percent, with log injection passing twelve percent and cross-site scripting passing fifteen percent.

Teams also gain actionable context engineering and memory management practices that standardize agent behavior across distributed repositories. Rather than repeatedly prompting agents with identical rules regarding package managers or directory structures, engineers consolidate verified instructions inside project configuration files like CLAUDE.md and AGENTS.md. The system also flags abandoned verification cycles, ensuring that whenever an agent fails a build or lint check, the diagnostic forces a rerun of that specific test after applying fixes. This systematic feedback loop prevents agents from prematurely declaring tasks complete before all validation criteria pass.

How to get Vibemetric – An open-source app that grades

Engineers can install the open-source release of Vibemetric directly onto macOS machines through the official download package distributed for Apple Silicon and Intel hardware. Because the application evaluates local transcript directories written by standard terminal tools, developers must ensure their CLI environments are running current versions of Anthropic Claude Code or OpenAI Codex. Updating the terminal client globally ensures full compatibility with the latest session telemetry formats and subagent definitions parsed by the desktop scoring engine:

npm install -g @anthropic-ai/claude-code
claude update

Once installed, developers configure runtime parameters and model routing rules directly inside active CLI sessions or system environment profiles to optimize cost tiers. Switching models dynamically during interactive work allows users to reserve frontier reasoning engines for complex architectural debugging while assigning routine tasks to lightweight models:

/model claude-fable-5-1
/model fable

To initialize a session pinned directly to a specific target model from the command line, launch the client with the model flag:

claude --model fable

Developers can permanently configure their default model routing by exporting the preferred engine variable within their shell configuration profile:

export ANTHROPIC_MODEL=claude-fable-5-1

What to watch after Vibemetric – An open-source app that grades

Following initial deployment, engineering leads must monitor whether agents consistently adhere to documented project guidelines or drift into legacy inefficiencies. A critical metric involves tracking the ratio of automated test executions against file modifications, ensuring agents do not revert to unverified code insertions. Teams must also inspect task planning compliance, particularly for multi-file refactoring efforts, where unstructured starts historically resulted in half of all initiatives requiring complete rebuilds from scratch. Maintaining strict pass and fail criteria before code generation begins prevents unnecessary rework and maintains code quality standards across active development branches.

Organizations should also track overall token consumption curves to verify that subagent routing rules remain effective as projects evolve. Workspaces that route documentation and email drafting to lightweight models like Haiku prevent routine copy tasks from consuming hundreds of thousands of tokens on top-tier engines. Regular review of session permission logs is equally essential to verify that agents operate under strict command sandboxes rather than broad root privileges. As session history accumulates, teams can use rolling daily scores to calibrate repository context files, refine team-wide skill libraries, and eliminate repetitive prompting friction permanently.