agent-flow-visualization
VS Code extension for real-time visualization of Claude Code agent orchestration as interactive node graphs
Best use case
agent-flow-visualization is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
VS Code extension for real-time visualization of Claude Code agent orchestration as interactive node graphs
Teams using agent-flow-visualization should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/agent-flow-visualization/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How agent-flow-visualization Compares
| Feature / Agent | agent-flow-visualization | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
VS Code extension for real-time visualization of Claude Code agent orchestration as interactive node graphs
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
Related Guides
SKILL.md Source
# Agent Flow
> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.
Agent Flow is a VS Code extension that provides real-time visualization of Claude Code agent orchestration. It renders agent execution as an interactive node graph, showing tool calls, branching, subagent coordination, and timing — turning Claude Code's black-box execution into a transparent, debuggable flow.
## Installation
### Via VS Code Marketplace
1. Open VS Code Extensions (`Cmd+Shift+X`)
2. Search for **Agent Flow** by simon-p
3. Click Install
Or install directly: [marketplace.visualstudio.com/items?itemName=simon-p.agent-flow](https://marketplace.visualstudio.com/items?itemName=simon-p.agent-flow)
### Requirements
- VS Code 1.85 or later
- Claude Code CLI installed and accessible
- Node.js (for Claude Code)
## Quick Start
```
# 1. Open Command Palette
Cmd+Shift+P (Mac) / Ctrl+Shift+P (Win/Linux)
# 2. Run the command
> Agent Flow: Open Agent Flow
# 3. Start a Claude Code session in your workspace
# Agent Flow auto-detects it and begins streaming
```
Or use the keyboard shortcut:
- **Mac**: `Cmd+Alt+A`
- **Win/Linux**: `Ctrl+Alt+A`
## Commands
| Command | Description |
|---------|-------------|
| `Agent Flow: Open Agent Flow` | Open visualizer in current editor column |
| `Agent Flow: Open Agent Flow to Side` | Open in a side editor column |
| `Agent Flow: Connect to Running Agent` | Manually connect to a specific agent session |
| `Agent Flow: Configure Claude Code Hooks` | Set up Claude Code hooks for live event streaming |
## Configuration
Settings available in VS Code settings (`settings.json`):
```jsonc
{
// Path to a JSONL event log file to watch/replay
"agentVisualizer.eventLogPath": "/path/to/agent-events.jsonl",
// Auto-open the visualizer when an agent session starts
"agentVisualizer.autoOpen": true,
// Development server port (0 = production mode, use built assets)
"agentVisualizer.devServerPort": 0
}
```
### Auto-Open on Agent Start
```jsonc
// settings.json
{
"agentVisualizer.autoOpen": true
}
```
## Claude Code Hooks Setup
Agent Flow uses Claude Code's hook system for zero-latency event streaming. Hooks are configured automatically on first open, but you can reconfigure manually.
### Automatic Configuration
Run from Command Palette:
```
> Agent Flow: Configure Claude Code Hooks
```
### Manual Hook Configuration
If you need to configure hooks manually, Agent Flow starts a local HTTP server that receives events. The hooks forward Claude Code lifecycle events (tool calls, responses, session start/end) to the extension.
Claude Code hooks are configured in `~/.claude/settings.json` or project-level `.claude/settings.json`:
```jsonc
{
"hooks": {
"PreToolUse": [
{
"matcher": ".*",
"hooks": [
{
"type": "command",
"command": "curl -s -X POST http://localhost:PORT/hook -H 'Content-Type: application/json' -d @-"
}
]
}
],
"PostToolUse": [
{
"matcher": ".*",
"hooks": [
{
"type": "command",
"command": "curl -s -X POST http://localhost:PORT/hook -H 'Content-Type: application/json' -d @-"
}
]
}
]
}
}
```
> Agent Flow manages this configuration automatically — prefer using the command palette command.
## JSONL Event Log Mode
For replaying past sessions or watching log files generated outside VS Code:
```jsonc
// settings.json
{
"agentVisualizer.eventLogPath": "${workspaceFolder}/logs/agent-session.jsonl"
}
```
Agent Flow tails the file and visualizes events as they arrive. Use this for:
- Post-hoc debugging of agent runs
- Sharing agent execution recordings with teammates
- CI/CD pipeline agent monitoring
### JSONL Event Format
Each line in the log file is a JSON event object:
```jsonl
{"type":"session_start","sessionId":"abc123","timestamp":"2026-03-21T10:00:00Z","model":"claude-opus-4-5"}
{"type":"tool_use","sessionId":"abc123","toolName":"bash","input":{"command":"ls -la"},"timestamp":"2026-03-21T10:00:01Z"}
{"type":"tool_result","sessionId":"abc123","toolName":"bash","output":"total 48\n...","timestamp":"2026-03-21T10:00:02Z"}
{"type":"message","sessionId":"abc123","role":"assistant","content":"I can see the files...","timestamp":"2026-03-21T10:00:03Z"}
```
## Multi-Session Support
Agent Flow tracks multiple concurrent Claude Code sessions with tabs. Each session gets its own visualization canvas.
```
# Start multiple Claude Code sessions in different terminals
# Each appears as a separate tab in Agent Flow
# Click tabs to switch between session graphs
```
## Interactive Canvas Features
### Navigation
- **Pan**: Click and drag on empty canvas
- **Zoom**: Scroll wheel / pinch gesture
- **Select node**: Click any agent or tool call node to inspect details
- **Reset view**: Double-click empty canvas
### Node Types in the Graph
| Node Type | Visual | Description |
|-----------|--------|-------------|
| Agent/Subagent | Circle | Claude instance making decisions |
| Tool Call | Rectangle | Individual tool invocation (bash, read_file, etc.) |
| Tool Result | Rectangle (dashed) | Output returned from tool |
| Branch | Diamond | Decision point spawning subagents |
| Return | Arrow | Subagent returning result to parent |
### Panels
- **Timeline**: Chronological view of all events with durations
- **Transcript**: Full message history between agent and tools
- **File Heatmap**: Which files received the most attention
## Contributing / Development Setup
```bash
git clone https://github.com/patoles/agent-flow
cd agent-flow
npm install
# Build the extension
npm run build
# Watch mode for development
npm run watch
# Run with dev server (hot reload)
# Set in settings: "agentVisualizer.devServerPort": 3000
npm run dev
```
### Extension Structure
```
agent-flow/
├── src/
│ ├── extension.ts # VS Code extension entry point
│ ├── hookServer.ts # HTTP server receiving Claude Code hook events
│ ├── sessionManager.ts # Manages multiple agent sessions
│ ├── webviewProvider.ts # Webview panel management
│ └── logWatcher.ts # JSONL file tail watcher
├── webview/
│ ├── src/
│ │ ├── App.tsx # Main React app
│ │ ├── Graph.tsx # Node graph canvas (likely D3 or React Flow)
│ │ ├── Timeline.tsx # Timeline panel
│ │ └── Transcript.tsx # Message transcript panel
│ └── package.json
└── package.json # Extension manifest
```
## Common Patterns
### Pattern: Debug a Failing Agent Run
1. Open Agent Flow (`Cmd+Alt+A`)
2. Run your Claude Code command that's failing
3. Watch the graph — look for:
- Red/error nodes indicating failed tool calls
- Unexpected branching
- Loops (agent retrying the same tool)
4. Click the failed tool node to inspect input/output
5. Check the Timeline panel for which tool call took unexpectedly long
### Pattern: Replay a Saved Session
```bash
# Save Claude Code output to JSONL during a run
# (depends on your Claude Code version/config)
claude --output-format jsonl > session-$(date +%s).jsonl
# Point Agent Flow at it
```
```jsonc
// .vscode/settings.json (project-level)
{
"agentVisualizer.eventLogPath": "./logs/session-1234567890.jsonl"
}
```
### Pattern: Auto-Visualize All Agent Sessions
```jsonc
// settings.json
{
"agentVisualizer.autoOpen": true
}
```
Now every time a Claude Code session starts in your workspace, Agent Flow opens automatically.
### Pattern: Side-by-Side Coding and Visualization
```
# Open Agent Flow to side so you can code and watch simultaneously
Cmd+Shift+P > Agent Flow: Open Agent Flow to Side
```
This opens the visualizer in a split editor, keeping your code files accessible in the main editor group.
## Troubleshooting
### Agent Flow doesn't detect my Claude Code session
1. Verify Claude Code is running in the same workspace folder
2. Check that hooks are configured: run `Agent Flow: Configure Claude Code Hooks`
3. Verify the hook server is running — look for "Agent Flow hook server listening" in the Output panel (`View > Output > Agent Flow`)
4. Try `Agent Flow: Connect to Running Agent` to manually specify the session
### Hooks aren't forwarding events
```bash
# Test the hook server manually
curl -X POST http://localhost:PORT/hook \
-H 'Content-Type: application/json' \
-d '{"type":"test","sessionId":"test123"}'
```
Check VS Code Output panel for the correct PORT value.
### Graph is empty / not updating
1. Check `agentVisualizer.eventLogPath` — if set, Agent Flow reads from file instead of hooks
2. Clear the setting if you want live hook-based streaming:
```jsonc
{ "agentVisualizer.eventLogPath": "" }
```
3. Reload VS Code window (`Cmd+Shift+P > Developer: Reload Window`)
### Extension not loading
```bash
# Check VS Code version meets requirement
code --version
# Must be 1.85 or later
# Check extension is enabled
# Extensions panel > search "Agent Flow" > verify enabled
```
### Multiple sessions showing in wrong tabs
Each session is identified by a unique session ID from Claude Code. If sessions are merging incorrectly, check that each `claude` process is started fresh (not reusing an existing session ID).
## Resources
- [VS Code Marketplace](https://marketplace.visualstudio.com/items?itemName=simon-p.agent-flow)
- [Demo Video](https://www.youtube.com/watch?v=Ud6eDrFN-TA)
- [GitHub Repository](https://github.com/patoles/agent-flow)
- [CraftMyGame](https://craftmygame.com) — the project that inspired Agent Flow
- [License: Apache 2.0](https://github.com/patoles/agent-flow/blob/main/LICENSE)Related Skills
ppt-image-first-workflow
Conversation-first, image-first PPT generation workflow skill using GPT Image 2 for full-page visual slides packaged into PPTX files.
mdv-markdown-visualization
MDV — Markdown superset for documents, dashboards, and slides with embedded charts, KPI stats, and data visualizations exported to HTML or PDF.
gstack-workflow-assistant
Team of specialist AI workflows for Claude Code with CEO review, engineering planning, code review, shipping, QA testing, and browser automation
git-city-3d-github-visualization
Build and extend Git City — a 3D pixel art city where GitHub profiles become interactive buildings using Next.js, Three.js, and Supabase.
geoflow-content-automation
GEOFlow open-source GEO/SEO content production system with AI generation, review workflow, and publishing pipeline built on PHP and PostgreSQL.
flowdriver-covert-transport
Tunnel SOCKS5 traffic through Google Drive API requests to bypass restrictive networks and DPI inspection.
fabro-workflow-factory
Skill for using Fabro, the open source AI coding workflow orchestrator that lets you define agent pipelines as Graphviz DOT graphs with human gates, multi-model routing, and cloud sandboxes.
```markdown
---
zeroboot-vm-sandbox
Sub-millisecond VM sandboxes for AI agents using copy-on-write KVM forking via Zeroboot
yourvpndead-vpn-detection
Android app that detects VPN/proxy servers (VLESS/xray/sing-box) via local SOCKS5 vulnerability, exposing exit IPs and server configs without root
xata-postgres-platform
Expert skill for Xata open-source cloud-native Postgres platform with copy-on-write branching, scale-to-zero, and Kubernetes deployment
x-mentor-skill-nuwa
AI-powered X (Twitter) content strategy skill that distills methodologies from 6 top creators + open-source algorithm data into actionable writing, growth, and monetization guidance.