Developers working with Anthropic’s Claude Code CLI often face a common frustration: opaque billing and mysterious behavior during long coding sessions. To solve this, an independent developer has launched Argus, a new VSCode extension that audits and debugs these AI interactions, according to a launch announcement on Hacker News.
Argus operates as a “time machine debugger” for the local .claude directory. It parses session data to explain exactly why a specific task cost $2 or why the AI read the same file a dozen times. This level of observability is becoming critical as developers move from simple chat interfaces to autonomous agentic workflows where token usage can spiral quickly.
Key Capabilities
The tool breaks down the “black box” of AI coding sessions through several distinct features:
- Step-by-Step Playback: Users can click through every step of a session to see specific tool inputs and outputs, identifying exactly where the agent went off track.
- Cost Analysis: Argus identifies which specific tools or actions burned the most tokens, allowing developers to optimize their workflows.
- Performance Insights: The extension highlights inefficiencies like retry loops, duplicate file reads, and context pressure.
- Visual Flow: It generates flow diagrams showing file dependencies and visualizes token usage, including cache hit rates.
Why It Matters
This launch highlights a growing need for tooling around AI agents. While the agents themselves are powerful, the infrastructure to monitor them is still immature. The developer noted that during their own testing, they discovered 40% of their session costs came from Claude Code repeatedly reading the same files. They also found that retry loops were costing an average of $0.15 each, costs that remain invisible without deep inspection tools.
Limitations and Availability
Argus is available now on the VSCode Marketplace and GitHub. However, the developer notes a few constraints. The tool currently only works with local directories, meaning it cannot debug remote sessions. Furthermore, because the Claude Code session format is undocumented, the extension relies on heuristic-based analysis that could break if Anthropic changes their data structure in future updates.
For developers relying on Claude Code for daily tasks, Argus offers a necessary layer of transparency to keep token budgets in check.