diff --git a/README.md b/README.md index c27050cd..bc9eb1e9 100644 --- a/README.md +++ b/README.md @@ -7,7 +7,7 @@

Terminal tells you nothing. This shows you everything.
- A desktop app that reconstructs exactly what Claude Code did — every file path, every tool call, every token — from the raw session logs already on your machine. + You're the CTO, agents are your team. They handle tasks themselves, message each other, review each other's code. You just look at the kanban board and drink coffee.

@@ -40,6 +40,7 @@
+ ## Installation @@ -64,103 +66,6 @@ The app reads session logs from `~/.claude/` — the data is already on your mac --- -## Why This Exists - -### Claude Code stopped telling you what it's doing. - -Recent Claude Code updates replaced detailed tool output with opaque summaries. `Read 3 files`. `Searched for 1 pattern`. `Edited 2 files`. No paths, no content, no line numbers. The context usage indicator became a three-segment progress bar with no breakdown. To get the details back, the only option is `--verbose` — which dumps raw JSON, internal system prompts, and thousands of lines of noise into your terminal. - -**There is no middle ground in the CLI.** You either see too little or too much. - -Claude Agent Teams UI restores the information that was taken away — structured, searchable, and without a single modification to Claude Code itself. It reads the raw session logs from `~/.claude/` and reconstructs the full execution trace: every file path that was read, every regex that was searched, every diff that was applied, every token that was consumed — organized into a visual interface you can actually reason about. - -### The wrapper problem. - -There are many GUI wrappers for Claude Code — Conductor, Craft Agents, Vibe Kanban, 1Code, ccswitch, and others. None of them solved the actual problem: - -**They wrap Claude Code.** They inject their own prompts, add their own abstractions, and change how Claude behaves. If you love the terminal — and I do — you don't want that. You want Claude Code exactly as it is. - -**They only show their own sessions.** Run something in the terminal? It doesn't exist in their UI. You can only see what was executed through *their* tool. The terminal and the GUI are two separate worlds. - -**You can't debug what went wrong.** A session failed — but why? The context filled up too fast — but what consumed it? A subagent spawned 5 child agents — but what did they do? Even in the terminal, scrolling back through a long session to reconstruct what happened is nearly impossible. - -**You can't monitor what matters.** Want to know when Claude reads `.env`? When a single tool call exceeds 4K tokens of context? When a teammate sends a shutdown request? You'd have to wire up hooks manually, every time, for every project. - -**Claude Agent Teams UI takes a different approach.** It doesn't wrap or modify Claude Code at all. It reads the session logs that already exist on your machine (`~/.claude/`) and turns them into a rich, interactive interface — regardless of whether the session ran in the terminal, in an IDE, or through another tool. - -> Zero configuration. No API keys. Works with every session you've ever run. - ---- - -## Key Features - -### :mag: Visible Context Reconstruction - -context - -Claude Code doesn't expose what's actually in the context window. Claude Agent Teams UI reverse-engineers it. - -The engine walks each turn of the session and reconstructs the full set of context injections — **CLAUDE.md files** (broken down by global, project, and directory-level), **skill activations**, **@-mentioned files**, **tool call inputs and outputs**, **extended thinking**, **team coordination overhead**, and **user prompt text**. - -The result is a per-turn breakdown of estimated token attribution across 7 categories, surfaced in three places: a **Context Badge** on each assistant response, a **Token Usage popover** with percentage breakdowns, and a dedicated **Session Context Panel**. - -### :chart_with_downwards_trend: Compaction Visualization - - - -**See the moment your context hits the limit.** - -When Claude Code hits its context limit, it silently compresses your conversation and continues. Most tools don't even notice this happened. - -Claude Agent Teams UI detects these compaction boundaries, measures the token delta before and after, and visualizes how your context fills, compresses, and refills over the course of a session. You can see exactly what was in the window at any point, and how the composition shifted after each compaction event. - - -### :bell: Custom Notification Triggers - - - -Define rules for when you want to receive **system notifications**. Match on regex patterns, assign colors, and filter your inbox by trigger. - -- **Built-in defaults**: `.env File Access Alert`, `Tool Result Error` (`is_error: true`), and `High Token Usage` (default: 8,000 total tokens). -- **Custom matching**: use regex against specific fields like `file_path`, `command`, `prompt`, `content`, `thinking`, or `text`. -- **Sensitive-file monitoring**: create alerts for `.env`, `secrets`, payment/billing/stripe paths, or any project-specific pattern. -- **Noise control**: choose input/output/total token thresholds, add ignore patterns, and scope triggers to selected repositories. - - -### :hammer_and_wrench: Rich Tool Call Inspector - -Every tool call is paired with its result in an expandable card. Specialized viewers render each tool natively: -- **Read** calls show syntax-highlighted code with line numbers -- **Edit** calls show inline diffs with added/removed highlighting -- **Bash** calls show command output -- **Subagent** calls show the full execution tree, expandable in-place - - -### :busts_in_silhouette: Team & Subagent Visualization - -Claude Code now spawns subagents via the Task tool and coordinates entire teams via `TeamCreate`, `SendMessage`, and `TaskUpdate`. In the terminal, all of this collapses into an unreadable stream. Claude Agent Teams UI untangles it. - -- **Subagent sessions** are resolved from Task tool calls and rendered as expandable inline cards — each with its own tool trace, token metrics, duration, and cost. Nested subagents (agents spawning agents) render as a recursive tree. -- **Teammate messages** — sent via `SendMessage` with color and summary metadata — are detected and rendered as distinct color-coded cards, separated from regular user messages. Each teammate is identified by name and assigned color. -- **Team lifecycle** is fully visible: `TeamCreate` initialization, `TaskCreate`/`TaskUpdate` coordination, `SendMessage` direct messages and broadcasts, shutdown requests and responses, and `TeamDelete` teardown. -- **Session summary** shows distinct teammate count separately from subagent count, so you can tell at a glance how many agents participated and how work was distributed. - -### :zap: Command Palette & Cross-Session Search - -Hit **Cmd+K** for a Spotlight-style command palette. Search across all sessions in a project — results show context snippets with highlighted keywords. Navigate directly to the exact message. - -### :globe_with_meridians: SSH Remote Sessions - -Connect to any remote machine over SSH and inspect Claude Code sessions running there — same interface, no compromise. - -Claude Agent Teams UI parses your `~/.ssh/config` for host aliases, supports agent forwarding, private keys, and password auth, then opens an SFTP channel to stream session logs from the remote `~/.claude/` directory. Each SSH host gets its own isolated service context with independent caches, file watchers, and parsers. Switching between local and remote workspaces is instant — the app snapshots your current state to IndexedDB before the switch and restores it when you return, tabs and all. - -### :bar_chart: Multi-Pane Layout - -Open multiple sessions side-by-side. Drag-and-drop tabs between panes, split views, and compare sessions in parallel — like a proper IDE for your AI conversations. - ---- - ## What the CLI Hides vs. What Claude Agent Teams UI Shows | What you see in the terminal | What Claude Agent Teams UI shows you | @@ -285,10 +190,6 @@ pnpm dist # macOS + Windows + Linux See [CONTRIBUTING.md](CONTRIBUTING.md) for development guidelines. Please read our [Code of Conduct](CODE_OF_CONDUCT.md). -## Acknowledgements - -Based on [claude-devtools](https://github.com/matt1398/claude-devtools) by matt1398. - ## Security IPC handlers validate all inputs with strict path containment checks. File reads are constrained to the project root and `~/.claude`. Sensitive credential paths are blocked. See [SECURITY.md](SECURITY.md) for details.