Krometrail
Gives AI coding agents eyes into running applications by recording browser activity and providing session investigation tools for debugging.
README
Krometrail
Browser observation and runtime debugging for AI coding agents.
curl -fsSL https://krometrail.dev/install.sh | sh
Krometrail is an MCP server and CLI that gives AI coding agents eyes into running applications. It records browser activity — network requests, console output, DOM mutations, framework state, storage changes, and screenshots — then lets agents search, inspect, and diff that recorded session to diagnose bugs. It also bridges the Debug Adapter Protocol (DAP) for breakpoint-level debugging across 6 languages.
Browser Observation
Record everything happening in a browser session and investigate it later — no code changes required.
What Gets Captured
- Network — every request/response with headers, bodies, status codes, timing, and WebSocket frames
- Console — all console output with levels, args, and stack traces
- DOM mutations — meaningful structural changes (forms, dialogs, sections)
- User input — clicks, form submissions, field changes
- Screenshots — periodic and navigation-triggered snapshots
- Browser storage — localStorage/sessionStorage mutations and cross-tab events
- Framework state — React and Vue component lifecycles, state/prop diffs, store mutations
- Framework errors — auto-detected anti-patterns (stale closures, infinite re-renders, missing cleanup)
Quick Start
Add to your agent's MCP config (.mcp.json in your project root):
{
"mcpServers": {
"krometrail": {
"command": "krometrail",
"args": ["--mcp"]
}
}
}
Start recording a browser session:
# MCP: chrome_start({ url: "http://localhost:3000", framework_state: true })
# CLI:
krometrail chrome start http://localhost:3000 --framework-state
# Place markers at significant moments
krometrail chrome mark "submitted form"
# Stop recording
krometrail chrome stop
Investigate what happened:
# List recorded sessions
krometrail session list --has-errors
# Get a structured overview
krometrail session overview <session-id>
# Search for specific events
krometrail session search <session-id> --event-types network_response --status-codes 500
krometrail session search <session-id> --framework react --pattern stale_closure
# Deep-dive into a specific event
krometrail session inspect <session-id> --event-id <id>
# Compare two moments (what changed between page load and error?)
krometrail session diff <session-id> --from <timestamp> --to <timestamp>
# Generate reproduction steps or test scaffolds
krometrail session replay-context <session-id> --format playwright
Browser MCP Tools
| Tool | Description |
|---|---|
chrome_start |
Launch Chrome and start recording (URL, framework observation, tab filtering) |
chrome_status |
Current recording status, event counts, active tabs |
chrome_mark |
Place a named marker in the recording timeline |
chrome_stop |
Stop recording and persist events to database |
session_list |
List recorded sessions with filters (time, URL, errors, markers) |
session_overview |
Structured overview: navigation, markers, errors, network, framework summary |
session_search |
Full-text and structured search across recorded events |
session_inspect |
Deep-dive into a specific event with full context and nearest screenshot |
session_diff |
Compare two moments: URL, storage, console, network, framework state changes |
session_replay_context |
Generate reproduction steps or Playwright/Cypress test scaffolds |
Framework State Observation
When enabled, Krometrail hooks into React DevTools and Vue Devtools to track:
- Component mount/update/unmount lifecycles
- State and prop changes with before/after diffs
- Render counts and trigger source identification
- Pinia/Vuex store mutations (Vue)
- Auto-detected bug patterns: stale closures, infinite re-renders, missing effect cleanup, excessive context re-renders
{ "url": "http://localhost:3000", "framework_state": true }
{ "url": "http://localhost:3000", "framework_state": ["react"] }
{ "url": "http://localhost:3000", "framework_state": ["react", "vue"] }
Runtime Debugging
Set breakpoints, step through code, and inspect variables across 6 languages via DAP.
Supported Languages
| Language | Debugger | Adapter | Status |
|---|---|---|---|
| Python | debugpy | TCP | Stable |
| Node.js | js-debug | TCP | Stable |
| Go | Delve | TCP | Stable |
| Rust | CodeLLDB | TCP | Stable |
| Java | java-debug-adapter | TCP | Stable |
| C/C++ | GDB 14+ / lldb-dap | stdio | Stable |
Debug CLI
krometrail launch "python app.py" --break order.py:147
krometrail step over
krometrail eval "discount"
krometrail vars --scope local
krometrail continue
krometrail stop
Debug MCP Tools
| Tool | Description |
|---|---|
debug_launch |
Launch a program with initial breakpoints |
debug_attach |
Attach to a running process |
debug_stop |
Terminate the debug session |
debug_status |
Query session state and capabilities |
debug_continue |
Resume execution until next breakpoint |
debug_step |
Step over, into, or out |
debug_run_to |
Run to a specific line |
debug_set_breakpoints |
Set breakpoints with conditions, hit counts, logpoints |
debug_set_exception_breakpoints |
Filter by exception type |
debug_list_breakpoints |
List all active breakpoints |
debug_evaluate |
Evaluate an expression in the current frame |
debug_variables |
Inspect variables by scope with regex filtering |
debug_stack_trace |
Get the full call stack |
debug_source |
Read source code around a location |
debug_watch |
Add/remove persistent watch expressions |
debug_action_log |
Review the investigation log |
debug_output |
Capture stdout/stderr |
debug_threads |
List threads, goroutines, etc. |
Features
- Passive browser recording — capture network, console, DOM, storage, and screenshots without code changes
- Framework-aware — React and Vue state tracking with bug pattern detection
- Session investigation — search, inspect, diff, and replay recorded browser sessions
- Compact viewport — debugger state rendered in ~400 tokens per stop, optimized for LLM context windows
- Conditional breakpoints —
order.py:147 when discount < 0, hit counts, logpoints - Watch expressions — persistent expressions auto-evaluated on every stop
- Framework detection — auto-detects pytest, Django, Flask, jest, mocha, go test
- Multi-threaded — thread/goroutine listing and selection
- Shell completions — tab-completion for bash, zsh, and fish via
krometrail completions
Skill File
Install the agent skill for CLI-based workflows. Install via skilltap or print to stdout:
skilltap install ./skill # Install via skilltap
krometrail skill # Or print skill to stdout
Development
bun install # Install dependencies
bun run dev # Run CLI in dev mode
bun run mcp # Run MCP server
bun run build # Compile single binary
bun run build:all # Build for all platforms (Linux, macOS, Windows)
Testing
bun run test # All tests
bun run test:unit # Unit tests (no external deps)
bun run test:integration # Integration tests (needs debuggers)
bun run test:e2e # E2E tests (full MCP path)
bun run test:agent # Agent harness scenarios
Integration and E2E tests require debuggers to be installed. Run krometrail doctor to check availability. Tests skip cleanly per-adapter when a debugger is not found.
Agent Harness
The agent harness (tests/agent-harness/) is a scenario-based test suite for evaluating how well agents debug with Krometrail. It contains 35 scenarios across 3 languages at 5 difficulty levels:
- Python — 12 scenarios (closure bugs, mutation errors, float accumulation, deep pipelines)
- Node.js — 11 scenarios (async races, event loop ordering, regex state,
thisbinding) - TypeScript — 12 scenarios (type assertion escapes, generic constraints, runtime registries)
bun run test:agent # Run scenarios
bun run test:agent:report # Generate report with token/cost metrics
Linting
bun run lint # Check with Biome
bun run lint:fix # Auto-fix
Architecture
src/
mcp/ MCP server + tool handlers
cli/ CLI entry point + commands (citty)
core/ Session manager, viewport renderer, DAP client, compression
adapters/ Language-specific debugger adapters (6 languages)
browser/ Chrome CDP recording, investigation engine, framework observers
daemon/ Session persistence over Unix socket
frameworks/ Auto-detection for test/web frameworks
The MCP server and CLI share the same core. Browser tools use CDP to record events into a SQLite-backed session store with JSONL event storage. Debug tools use the session manager to orchestrate DAP communication. The viewport renderer formats state for agents, and adapters handle language-specific debugger setup.
Documentation
| Document | Contents |
|---|---|
| VISION.md | Problem statement, prior art, roadmap |
| ARCH.md | System layers, data flow, viewport rendering |
| UX.md | Viewport abstraction, agent interaction patterns |
| SPEC.md | Adapter contract, type definitions |
| INTERFACE.md | MCP tool + CLI command reference |
| TESTING.md | Testing philosophy and tiers |
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.