## What is Thermite?

Thermite is a self-hosted error tracker that speaks Sentry's wire protocol. Point an unmodified
Sentry SDK at it and events group into issues in your own PostgreSQL — no Sentry account, no error
data leaving your infrastructure.

What makes it different is who gets to act on the data. Everything the dashboard shows — the
exception chain, every stack frame with source context, breadcrumbs, tags — is also served over an
MCP server, shaped so a coding agent can diagnose a bug in one request instead of walking a
resource tree. Claude Code connects with an API key; claude.ai connects through the built-in OAuth
flow. Both hit the same tools.

> **Tip:** One binary, one Postgres. No Redis, no message broker, no worker pool.

## The triage loop

Thermite never calls a model. It hands the work to your own agent:

1. **An exception lands.** Ingest groups it into an issue and — in the same transaction — queues
   exactly one unit of triage work, however large the error storm behind it.
2. **Your agent claims the work over MCP.** The claim is leased, so two agents never triage the
   same bug twice, and an agent that dies releases its work. The item carries the release that
   crashed — when that is a git SHA, the agent can check out the revision that actually broke.
3. **The analysis comes back.** Root cause and suggested fix land on the issue page, next to the
   alert that told you a human should look.

## Beyond errors

**Cron monitoring**
Jobs check in on their schedule through the same DSN. A missed or overrunning run becomes an
ordinary error event — grouped, alerted, and triaged like any other.

**Release health**
Crash-free rate per release, counted from SDK sessions, so a busy release does not read as a
broken one.

**Alerts per project**
Email and webhook routing per project, delivered at-least-once from an outbox. A duplicate is
possible; a silently dropped alert is not.

**Component keys**
Mint labeled DSNs for the parts of one product — 'worker', 'saas' — and filter by component
without splitting the issue stream into separate projects.

## Next steps

**Self-hosting**
Run the Docker image: one compose file, one Postgres, no identity provider.

**Development setup**
Build it from source and send it a first event.

**Connect an agent**
Wire Claude Code or claude.ai to the triage queue over MCP.

**Authentication**
API keys for agents, and how people sign in.