Let Claude Code or Cursor read your production errors

An error comes in while you are mid-feature. The usual move is to open the issue in a browser, copy the minified frames, paste them into Claude Code, and explain which release is live. Epure removes the copy step: the agent asks for the issue itself and gets the stack, the release and the count in one answer.

How does Claude Code read an Epure issue?

Claude Code reads an Epure issue through the epure_issues_context MCP tool. One call returns an issue brief: the error, its source-mapped stack when you uploaded source maps, the release and the occurrence count. Raw payloads and end-user data stay out of the brief unless the agent asks for them explicitly. Ask "explain the latest Epure issue" and the agent lists issues with epure_issues_list, then fetches the brief.

# Epure Cloud: hosted MCP with OAuth, no local install
claude mcp add --transport http --scope user epure https://app.epure.sh/mcp

# Any instance, including self-hosted: local stdio server
npx -y @epure-sh/mcp login
claude mcp add epure -- npx -y @epure-sh/mcp

For Cursor, npx -y @epure-sh/mcp init prints a config with no credential in it. The step-by-step setup, with the first call to check, is in connect Epure to Claude Code, Cursor or Codex.

Is it safe to let a coding agent read production errors?

Epure marks event data as untrusted before your agent reads it. Error messages and breadcrumbs come from your users' traffic, so a message can contain text written to steer an agent. The Epure MCP server's instructions tell the agent that application content is data, never instructions, and epure_issues_context leaves raw payloads out and hashes end-user emails and IDs unless the call asks for them. Epure also never stores IP addresses. Keep write scopes away from unattended agent loops.

Scope the access to the task. Sign in with --scope read:agent for read-only consent, or start the server with --read-only so only read tools are listed. Resolving or snoozing needs write:triage, and destructive actions require an explicit confirm: true. Workspace roles still apply on the server, so a token never grants more than its user has. The trade-offs are covered in giving your coding agent production errors without trusting the payload.

Does agent access work with self-hosted Epure?

Yes. Self-hosted Epure, free under AGPL-3.0, uses the same MCP server and CLI. Add --url https://errors.example.com to the login and server commands; HTTPS is required except on loopback during development. The hosted MCP endpoint at /mcp runs on Epure Cloud only. Credentials stay in a local OAuth store shared by the CLI and the MCP server, never in the MCP config file.

What does Epure not do for your coding agent?

Epure gives the agent evidence, not a fix. It does not write patches, open pull requests or run a debugging agent of its own; Claude Code, Cursor or Codex does the reasoning, and you review the diff and run the tests. Sentry goes further here: it offers a hosted MCP server and Seer, its debugging agent, at $40 per active contributor per month (Sentry pricing, checked 6 November 2026). Pick Sentry if you want the vendor's agent; pick Epure if your own agent is enough.

Which Epure plans include MCP and the CLI?

Every Epure plan includes MCP and the CLI, with no per-seat or per-contributor charge. Cloud Pro is $29 a month for 200,000 events, and the 14-day trial runs on Pro limits with no card. Self-hosting is free. Exact options and scopes are in the MCP and CLI reference; plan limits are on the Epure pricing page.