DEV Community

Gentrit Biba
Gentrit Biba

Posted on

Claude Code Is My Favorite Dev Tool — I Was Flying Blind Until I Found Cogpit

I've been using Claude Code daily for months. It writes code, runs tests, manages git branches, spawns sub-agents — the whole deal. On a good day, I'll have multiple sessions running across different projects, background agents churning through tasks, and a team of sub-agents coordinating work.

The problem? All of this happens inside a terminal. A single, scrolling terminal.

You don't know how many tokens you've burned until you check the billing page. You can't see what a background agent decided to do three minutes ago without scrolling through walls of JSONL. And if you're running six agents in parallel? Good luck keeping track.

I started looking for something better. That's when I came across Cogpit.

What Cogpit Actually Is

Cogpit is a real-time dashboard that reads your Claude Code session files and turns them into something you can actually navigate. It's a desktop app (macOS and Linux) or a browser-based dev server — no API keys, no separate login. It uses your existing Claude CLI and reads the JSONL files that Claude Code already writes to ~/.claude/projects/.

You install it, open it, and immediately see every project and session you've ever run. Live sessions show up with status indicators — green for running, blue for tool use, amber for thinking. The kind of stuff that's invisible in a terminal.

The Session Timeline Changes Everything

This is where Cogpit clicked for me. Every session renders as a structured conversation timeline:

  • User messages and assistant responses with rendered Markdown and syntax highlighting
  • Thinking blocks — expandable, so you can see Claude's reasoning without it cluttering the view
  • Tool calls with color-coded badges — Read is blue, Write is green, Edit is amber, Bash is red. Each one expandable with full input/output
  • Edit diffs — actual line-by-line diffs with syntax highlighting, not raw JSON
  • Sub-agent activity — color-coded panels showing what each spawned agent is doing

There's also a full-text search that filters across everything — user messages, assistant text, thinking blocks, tool names, tool output. Case-insensitive, real-time.

When you scroll past a turn's prompt, a sticky header keeps showing it so you always know which turn you're reading. Small thing, massive quality-of-life improvement when you're debugging a 40-turn session.

Live Streaming and Chat

Cogpit connects to active sessions via Server-Sent Events. New turns appear as Claude works — no refreshing, no polling on your end. The UI coalesces rapid updates so it stays smooth even when the agent is making 10 tool calls in a row.

You can also send messages to running sessions directly from Cogpit. Pick a model override per message (Opus, Sonnet, Haiku), use slash commands with autocomplete, drag-and-drop images, or even use voice input powered by Whisper running locally in WASM. There's a plan approval bar when the agent enters plan mode, and a question bar when it asks for input.

Basically, I stopped switching back to the terminal.

Token Analytics and Cost Tracking

This one caught me off guard. Cogpit breaks down every session with:

  • Per-turn token usage — input, output, cache creation, cache read
  • Model-aware cost calculation across Opus, Sonnet, and Haiku
  • A visual bar chart of token usage per turn
  • Context window usage as a percentage with color coding (green to red)
  • API rate limit tracking — 5-hour and 7-day utilization with time-to-reset

I had a session that burned through $7.57 in 10 turns. Cogpit showed me exactly where: most of it was cache creation from large file reads in the early turns. Without this visibility, I would've kept repeating the same pattern.

Undo/Redo with Branching

This feature alone is worth the install. You can rewind any session to a previous turn, and Cogpit handles the branching:

  • Create branches from any point in the conversation
  • Switch between branches via an SVG graph visualization
  • File operations (Edit/Write) are reversed on undo and replayed on redo
  • Ghost turns show archived content with hover-to-redo

I've used this to backtrack when an agent went down a wrong path — instead of starting over, I just branch from the last good turn and try a different prompt.

Teams and Multi-Agent Workflows

If you use Claude Code's team features (multiple agents coordinating via shared task lists), Cogpit gives you a dedicated dashboard:

  • Member cards with status indicators
  • A kanban-style task board (pending, in-progress, completed)
  • Color-coded message timeline showing inter-agent communication
  • Jump directly to any team member's session

Running a team of agents without this is like managing a distributed team over text files. Cogpit makes it visual.

The Details That Add Up

A few smaller things that show this was built by someone who actually uses Claude Code daily:

  • MRU session switching — Ctrl+Tab cycles through recently used sessions, Firefox-style
  • Process monitor — lists all system-wide claude processes with PID, memory, and CPU usage; detects orphaned processes
  • Worktree management — list active worktrees, see dirty/clean status, create PRs directly, bulk cleanup stale ones
  • File changes panel — tracks every Edit and Write operation in a session with +/- line counts
  • Todo progress tracking — when Claude uses TodoWrite, Cogpit shows a sticky progress bar with completion percentage and currently active task
  • Three themes — dark (default), deep OLED, and light. The dark theme uses OKLCH color space with hue-matched shadows. Not just "dark gray background."
  • Network access — enable it in settings, set a password, and access Cogpit from your phone on the same LAN. Full functionality, not a stripped-down mobile view
  • Terminal integration — open a terminal at any session's working directory; auto-detects Ghostty, iTerm, Warp, Alacritty, kitty

Who Is This For

If you use Claude Code casually — one session at a time, quick questions — you probably don't need this. The terminal is fine.

But if you're running long sessions, spawning background agents, managing costs across multiple projects, or using team workflows — Cogpit fills a gap that the CLI simply doesn't cover. It's the difference between flying a plane with instruments and flying by feel.

Try It

Cogpit is free and open source. Desktop app for macOS and Linux, or run it in the browser:

cogpit.gentrit.dev

You need Claude Code installed — Cogpit reads your existing sessions, it doesn't replace anything. Install it, open it, and you'll immediately see every session you've ever run.

I write about dev tools and performance on my blog. If you found this useful, check it out.

Top comments (0)