The story of Claude Code becomes an observable system, but not without its nuances
Sh0ny
10 августа 20262 min read
In short
The Claude Code History Viewer aggregates scattered JSONL logs into an interface with search, statistics, and an API. But when deployed to a server, it becomes clear that this isn’t just a viewer—it’s a standalone service that requires resources and careful handling of the logs.
Claude Code's history is typically stored in machine-readable JSONL, which is difficult to read and analyze. The Claude Code History Viewer transforms these records into a dashboard: you can reconstruct dialogs, view tool and subagent calls, find the specific snippet you need, and assess token consumption and cost.
The main value here isn’t in a pretty interface. The session history becomes almost like an observability log for AI development: it lets you understand exactly what the agent was doing, where additional executors were connected, and how costs escalated. This is especially useful if you’ve accumulated a large number of such sessions.
The viewer works with more than just Claude Code. It supports session histories for 27 other clients, including Codex CLI and Cursor. The same operations can be invoked via HTTP—a POST request to /api/<name>—so the tool can potentially be integrated into your own internal workflow rather than having to open it manually each time.
To run it on a server without a graphical interface, a headless mode is provided in Docker: a ready-to-use binary and a lightweight image. But this is exactly the kind of situation where the word “set up” sounds easier than the actual operation. It’s already interesting to view a 3.4 GB log file in the Global Overview, but the official Compose file hits the 512 MB memory limit. Additionally, the logs contain minor issues such as truncated tokens.
Bottom line: This isn’t a universal analytics platform, nor is it a magic window into the agent’s operations. It’s a convenient layer on top of local logs that can provide structure and statistics, but when running on a server, it requires checking memory, data format, and access to the interface. The longer the history and the more clients you have, the more important it is to view it as a data processing service rather than a simple file viewer.
Would you use such a dashboard to analyze AI agent errors, or would you store its session history in a separate service instead? Source: All Articles / Artificial Intelligence / Habr