Runheed

npx add-mcp 'https://runheed.com/api/mcp'

An MCP server for your production errors

Runheed collects the exceptions your apps throw in production. Its MCP server hands them to your coding agent: Claude Code, Cursor or any MCP client can search them, read the stack trace and the steps before the failure, fix the code and mark the error resolved.

The tools

Tool	What it does	Ask your agent
runheed_get_recent	The newest errors across your projects.	"What broke in production today?"
runheed_search_errors	Search error groups by text, status (unresolved, resolved, ignored) and project.	"Find unresolved TypeErrors in the checkout project."
runheed_get_error	One error in full: stack trace, breadcrumbs, agent context and recent events. If it came back after a fix, the note it was resolved with.	"Show me the stack trace for the top error."
runheed_get_error_context	Everything recorded around an error: all breadcrumbs, environment, release, request, user and agent session.	"Which agent session shipped the code behind this error?"
runheed_get_trends	Error counts and top offenders over 24 hours, 7 days or 30 days.	"Did errors go up after yesterday's deploy?"
runheed_get_project_health	A 0 to 100 health score, plus which error to fix first (by ID) and which ones came back after a fix.	"What should I fix first after this deploy?"
runheed_resolve_error	Mark an error resolved with a note such as the fixing commit. It reopens on its own if it happens again.	"Fix it and resolve it with the commit hash."
runheed_ignore_error	Ignore an expected error so it stops alerting.	"Ignore the bot 404s on /wp-login.php."
runheed_unignore_error	Stop ignoring an error.	"Start alerting on that timeout again."
runheed_list_projects	Your projects, each with its DSN.	"Which projects do I have, and what is the DSN for shop-api?"
runheed_create_project	Create a project and get its DSN, so your agent can set up error reporting in a new app without the dashboard.	"Set up Runheed in this repo."
runheed_send_test_event	Send one test error through the real ingest path to check the setup works end to end.	"Check that error reporting works."

Add it to your agent

The server runs at one URL. Add it to your client and sign in to Runheed when it asks; there is no key to copy and nothing to install.

https://runheed.com/api/mcp

Add to claude.ai Add to Cursor Add to VS Code Add to Replit

In claude.ai or Claude Desktop, open Runheed in Claude's connector directory and click Connect (or add it under Customize > Connectors > Add custom connector). In ChatGPT, add it under Plugins > + > Add custom MCP server. In Claude Code:

claude mcp add --transport http runheed https://runheed.com/api/mcp

Cursor and other clients that take a URL:

{
  "mcpServers": {
    "runheed": {
      "url": "https://runheed.com/api/mcp"
    }
  }
}

Or run it locally with an agent key

For clients without remote servers, or a machine where a sign-in window can't open: create an agent key under Settings > Agent keys and start the local server with npx.

claude mcp add runheed --env RUNHEED_API_KEY=al_your_key -- npx -y https://runheed.com/dl/runheed-mcp-0.1.7.tgz

{
  "mcpServers": {
    "runheed": {
      "command": "npx",
      "args": [
        "-y",
        "https://runheed.com/dl/runheed-mcp-0.1.7.tgz"
      ],
      "env": {
        "RUNHEED_API_KEY": "al_your_key"
      }
    }
  }
}

The npm and PyPI releases are not out yet, so these commands install the same package files straight from runheed.com. Nothing else changes when the registry versions ship.

Exact steps per client: claude.ai · ChatGPT · Claude Code · Cursor · Claude Desktop · VS Code · Windsurf · Codex CLI · Gemini CLI · Kiro · Lovable · Replit · Bolt · Antigravity · Zed · Cline

Send errors from your app

The MCP server reads what your apps report. Setup for your stack: Next.js · Node.js and Express · Python and FastAPI · Django · Flask · Supabase Edge Functions · Fastify · Vercel Functions · AWS Lambda (Node.js and Python) · Cloudflare Workers · Bun (Bun.serve) · Deno Deploy · Render (Node.js and Python) · Railway (Node.js and Python) · Go · Ruby · PHP · .NET and ASP.NET Core · Java and Spring Boot · React and browser JavaScript. Anything else can post JSON to the ingest endpoint (docs).

Put it to work

Start with the ways to give a coding agent your production errors. Then make it routine: check errors at the start of each Claude Code session, read a morning digest or file GitHub issues from new errors. Weighing other servers? See error tracking MCP servers compared. Building your own server? See how to track errors in an MCP server.

Questions

Is it a local or a remote MCP server?

Both. The remote server at https://runheed.com/api/mcp needs no install and no key to copy: add the URL to claude.ai, ChatGPT, Claude Desktop, Claude Code, Cursor, VS Code or Codex and sign in when it asks (OAuth). The local server runs with npx and an agent key, for Windsurf, Gemini CLI, Zed, Cline or any machine where a sign-in window can't open. Both have the same nine tools.

What can a connected app or agent key do?

Read errors, trends and project health in your workspace, and resolve, ignore or unignore errors. It cannot change billing, projects or other keys. Each app that signs in gets its own key, listed under Connected apps in Settings; keys you create start with al_, are stored as a hash and are shown once. Revoke either one in Settings at any time.

Where do the errors come from?

Your app sends them with the Node.js or Python SDK, or by posting JSON to the ingest endpoint. The MCP server only reads what your apps have reported.

Does it work alongside Sentry?

Yes. Keep Sentry for tracing and dashboards and add Runheed for the agent loop. Both SDKs run in the same process.

Try it on sample data 7-day trial, then $29/mo The MCP server, connected apps and unlimited agent keys come with every plan