Agents and MCP
Let your coding agent read and fix your production errors
Errilo has a remote MCP server. Your agent (Claude Code, Cursor, VS Code or an agent in CI) reads the issue, gets a stack trace that points to the right file at the right commit, fixes the code and marks the issue resolved. Included in every plan, the free one too. Errilo runs no AI model.
Errilo is in development. The MCP server runs at app.errilo.com/mcp; each connection method below says whether we have tested it.
What your agent gets
- The stack trace, translated with your source maps: your application’s frames first, with paths relative to your repository (
webpack://or/app/prefixes removed) andnode_modulesframes folded. - The commit of the release, when the release records one (sent with
sentry-cli releases set-commits, or a release version that is a commit SHA): the agent opens the right file at the right version. - The context: first and last occurrence, count, releases affected, regression, latest breadcrumbs, and which browsers, versions or URLs are hit.
- Compact Markdown rather than the raw event JSON: fewer tokens for the same information.
Tools
Eight tools, six to read and two to write:
| Tool | Scope | What it does |
|---|---|---|
list_projects | issues:read | Lists the projects the agent can see. |
search_issues | issues:read | Searches issues with filters: project, status, environment, release, text in the title. |
get_issue | issues:read | Reads an issue: status, occurrences, releases, stack trace and latest event. |
get_event | issues:read | Reads another occurrence of the issue. |
get_tag_distribution | issues:read | Shows which browsers, versions, environments or URLs the issue hits. |
get_release | issues:read | Shows what a release contains: its commit, new issues and regressions. |
update_issue | issues:write | Resolves an issue, resolves it in the next release, ignores it, assigns or unassigns it. |
add_comment | issues:write | Comments on an issue, for example with the diagnosis and the link to the pull request. |
And a fix_issue prompt that gives the agent a method: read the issue, check out the release commit, reproduce the error with a failing test, fix it, open a pull request (never a push to the main branch), comment, then mark the issue “resolved in the next release”. If the error comes back after the deploy, the issue regresses and you get an alert again.
No tool deletes anything. Search works with filters, not natural language: your agent turns your request into filters itself.
Connect from your editor
Add the server URL to your client. The first time, it opens Errilo in your browser: you sign in, choose the organization and the scopes, issues:read to read (the default) and issues:write to resolve, assign and comment. There is no token to copy.
claude mcp add --transport http errilo https://app.errilo.com/mcpThen type /mcp in Claude Code to sign in.
In Cursor or VS Code, one click adds the server, then you sign in when the editor asks:
Add to Cursor Install in VS Code
Or add it to the configuration file by hand:
{
"mcpServers": {
"errilo": { "url": "https://app.errilo.com/mcp" }
}
}{
"servers": {
"errilo": { "type": "http", "url": "https://app.errilo.com/mcp" }
}
}In VS Code, the MCP: Add Server command does the same. On claude.ai, open Settings, Connectors, add a custom connector and enter https://app.errilo.com/mcp.
Tested on 18 November 2026, from sign-in to resolving an issue: Claude Code with the /mcp sign-in, the claude.ai custom connector and MCP Inspector. Cursor and VS Code use the same MCP standard; their setups follow each client’s documentation, we haven’t tested them ourselves.
To cut an agent off, open Settings, Connected applications in Errilo and revoke its access: it stops working at once.
An agent in CI: a project token
An agent without a browser uses an access token restricted to one project. In Errilo, open Settings, API tokens, New token, tick issues:read (and issues:write to resolve and comment), and choose the project. The MCP server refuses tokens that aren’t restricted to a project.
# ERRILO_TOKEN comes from your CI secrets
claude mcp add --transport http errilo https://app.errilo.com/mcp \
--header "Authorization: Bearer $ERRILO_TOKEN"The token (errilo_…) is shown only once. Keep it in your CI secrets or an environment variable, never in a committed file.
Tested on 18 November 2026 with Claude Code: reading, resolving and the fix_issue prompt.
To start the agent when an error appears, an Errilo alert can trigger a GitHub workflow. The GitHub alert channel sends a repository_dispatch event of type errilo-alert, once per issue, with the issue ID:
on:
repository_dispatch:
types: [errilo-alert]
jobs:
fix:
runs-on: ubuntu-latest
steps:
- env:
ISSUE_ID: ${{ github.event.client_payload.issue_id }}
run: echo "Errilo issue $ISSUE_ID"The event also carries text from your errors, such as the title. Treat it as data: pass it through an environment variable, as above, and never write it straight into a shell command or an agent’s instructions.
A complete workflow, from the alert to the pull request, with its safeguards: auto-fix with Claude Code and GitHub Actions, step by step.
What keeps it safe
- Error messages, breadcrumbs and tags come from your users, so possibly from an attacker. Errilo hands them to the agent cleaned, truncated and wrapped in tags that mark them as untrusted data, after personal data has been scrubbed. The
fix_issueprompt tells the agent never to follow instructions found there. - Read-only by default. Writing needs
issues:write, and no tool deletes anything. - Every change to an issue, from the web app, a token or MCP, goes to the organization’s audit log with the client that made it. Owners and admins can read it.
- Limits per token: 120 requests a minute and 60 writes an hour.
Included in every plan
The MCP server is included in every plan, the free one too. Errilo runs no AI model: the agent you already pay for (Claude Code, Cursor…) does the work, and your error data goes to it, nowhere else. There is nothing for us to bill.
Comparing with Sentry’s AI agent? Read Errilo as a Sentry Seer alternative. Still on Sentry? The migration guide takes one line.