Broccoli-AI-Debugger
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
π₯¦ Broccoli β AI Debugger for VS Code
Broccoli turns "the agent reads a stack trace and guesses" into "the agent drives the real debugger": it sets breakpoints, steps, inspects live variables over the Debug Adapter Protocol, and proposes a fix you approve in a diff β never applied silently.
Two ways to use it:
Built-in agent β bring your own API key (Anthropic, OpenAI, Groq, DeepSeek, Together, xAI, Ollama, or any OpenAI-compatible endpoint) and debug from the sidebar.
MCP server β expose the same debugger tools over a localhost Model Context Protocol endpoint so external agents like Claude Code or Cursor can debug through your editor.
Quick start (built-in agent)
Open the Broccoli icon in the activity bar.
Click configure and pick a provider + model (the key is stored in VS Code's SecretStorage, never on disk).
Make sure your workspace has a
launch.jsondebug configuration.Describe the bug in the panel and hit Debug it.
You'll see the agent's activity live: every tool call as a trace line, its reasoning as cards, and token usage at the bottom. Any code change opens a diff with a modal Apply / Reject β nothing is written without you.
Related MCP server: agnt
MCP server (for Claude Code, Cursor, β¦)
Run "Broccoli: Start MCP Server" from the command palette (or set broccoli.mcp.autoStart). Then register it with your agent:
# Claude Code
claude mcp add --transport http broccoli http://127.0.0.1:4923/mcp// Cursor (~/.cursor/mcp.json)
{
"mcpServers": {
"broccoli": { "url": "http://127.0.0.1:4923/mcp" }
}
}The agent gets 16 tools: session lifecycle (start_debug_session by launch-config name, restart, stop), breakpoints (incl. conditional), stepping (continue_execution, step_over/into/out β each blocks until the next stop and returns the stop reason, top frame and a locals preview), inspection (get_stack_trace, inspect_variables, expand_variable), read_source, get_debugger_state for cheap orientation, and propose_code_fix β which still shows you the modal diff.
Settings:
Setting | Default | Purpose |
|
| Localhost port (server binds |
| (empty) | Optional |
|
| Start with the extension |
|
| Turn cap per built-in agent run |
|
| Token cap per run (0 = off) |
Safety model
These are properties of the implementation, not the prompt:
No expression evaluation β the agent cannot run code in the debuggee.
No silent edits β every change goes through a modal diff; line numbers are validated against the real file first; if the file changed while you reviewed, the apply is auto-rejected.
Workspace-scoped reads β
read_sourceandpropose_code_fixrefuse paths outside open workspace folders.Validated input β every tool call (internal or MCP) is checked against its JSON Schema before touching the debugger.
Bounded runs β turn cap, token budget, cancellation, and per-request timeouts.
Localhost-only MCP β bound to
127.0.0.1, optional bearer auth, DNS-rebinding protection.
Development
npm install
npm run compile # typecheck + esbuild bundle
npm test # unit + integration (real js-debug sessions against test-fixtures/)test-fixtures/ contains small buggy Node programs (off-by-one, wrong accumulator, async race, uncaught throw) used by the integration tests β and handy for trying the agent yourself: open test-fixtures/ as a workspace and ask Broccoli why sum is NaN.
See architecture.md for the full design.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelogβ¦
This server cannot be deployed
Maintenance
Related MCP Connectors
Live browser debugging for AI assistants β DOM, console, network via MCP.
Watches your live app, explains every problem in plain language and writes the fix prompt.
Browser-backed QA with evidence and fix-ready reports for coding agents.
Finds what's broken in your product, explains why, and fixes it from your agent.
Related MCP Servers
- AlicenseAqualityAmaintenanceEnables AI agents to perform step-through debugging of Python, JavaScript/Node.js, and Rust programs using the Debug Adapter Protocol, with support for breakpoints, variable inspection, and stack traces.29194MIT
- AlicenseNot gradedqualityBmaintenanceBridges AI coding agents with the browser to provide visual debugging, real-time error capture, screenshot capabilities, DOM inspection, and interactive wireframing through a reverse proxy with injected developer tools.22Apache 2.0
- AlicenseNot gradedqualityDmaintenanceBringing the debugging we know and love as human programmers to our AI agents β debug any supported language with breakpoints, variable/state inspection, and stepping, to supercharge agents' capabilities to reason about runtime code.44 PyPI22Apache 2.0
- AlicenseAqualityDmaintenanceEnables LLMs to automatically diagnose coding errors through codebase search, test execution, and live debugger integration (DAP/V8 CDP). Provides a secure, policy-gated environment for investigating failures while preventing destructive operations.9ISC