Detailed Analysis
A developer has created Conan, a native macOS application designed to provide real-time observability into Claude Code sessions — addressing a fundamental usability gap in Anthropic's terminal-based AI coding assistant. Claude Code, which operates through a command-line interface, gives users little to no visibility into its internal operations mid-session: which tools are being invoked, how much context window has been consumed, or what the session is costing in real time. Conan positions itself as a heads-up display (HUD) that runs alongside Claude Code, reading from the local `~/.claude` session data directory on-device and rendering that information into a live, structured interface. Core features include a streaming timeline of prompts, tool calls, and skills; a context-window fill gauge; a usage and cost monitor; and a skills and MCP (Model Context Protocol) browser.
The most technically distinctive feature of Conan is its skill-selection transparency. Claude Code internally selects from a set of skills or strategies each turn, and until now users had no insight into which skill was chosen or which alternatives were considered. Conan surfaces this decision layer, effectively exposing a slice of Claude's reasoning process that the terminal interface entirely obscures. This kind of interpretability — even at the shallow level of tool and skill selection — has practical value for power users who are trying to understand why Claude Code behaved a certain way, debug unexpected outputs, or optimize their prompting strategies to steer toward more effective skill usage.
The privacy architecture of Conan is worth noting as a deliberate design choice. All processing occurs on-device, with no telemetry and no data leaving the user's machine. This matters in a landscape where developer tools that integrate with AI systems often require cloud connectivity or API hooks, creating potential data exposure concerns — especially for users working with proprietary codebases. By reading only local session files and rendering everything natively on macOS, Conan sidesteps these concerns entirely, which also explains why the author chose to emphasize it prominently. The architecture also implies that the tool is inherently limited to macOS and to whatever Claude Code writes to its local session files, meaning its feature surface is bounded by what Anthropic exposes through that data format.
Conan represents a growing category of third-party tooling emerging around Claude Code and AI coding assistants more broadly — tools that don't modify the underlying model or its behavior but instead wrap the developer experience layer around it. This mirrors patterns seen historically with other developer platforms: when a core tool becomes widely adopted but lacks observability, a secondary ecosystem of dashboards, monitors, and debuggers emerges to fill the gap. The context-window gauge is particularly telling, as context exhaustion is one of the most disruptive and least visible failure modes in long Claude Code sessions, capable of silently degrading output quality before the user realizes what has happened. Cost monitoring similarly addresses a real pain point, as Claude Code usage can accumulate API charges rapidly during intensive sessions.
The broader significance of Conan lies in what it reveals about the maturation of Claude Code as a professional tool. Developer demand for this level of observability — session timelines, skill introspection, cost telemetry — signals that Claude Code has moved beyond casual experimentation into daily professional workflows where accountability and predictability matter. Anthropic has not yet natively surfaced this information in Claude Code's interface, leaving a gap that independent developers are now filling. Whether Anthropic eventually incorporates similar observability features natively, or whether the third-party ecosystem around Claude Code continues to expand, the existence of Conan illustrates that serious users of AI coding assistants are no longer satisfied with black-box terminal interactions and are actively building the infrastructure to understand what their AI tools are actually doing.
Read original article →