Condux error monitoring MCP server

Condux

Condux is a source-available error monitoring platform with a built-in AI fix engine.

SDKs report errors to a stateless ingest relay, which authenticates the DSN, applies per-plan rate limits and quotas, and scrubs PII before buffering through Redpanda. A consumer fingerprints and groups events into issues, writes time series to ClickHouse and issue state to PostgreSQL. A Next.js dashboard reads that back as grouped issues with stack traces, breadcrumbs, tag facets and per-hour charts, and alerting fires on new and regressed issues to email, Slack, Discord and webhooks.

The Conductor

The fix engine takes an issue to a draft pull request. It assembles a scoped, scrubbed context from the stack trace and the project's code mappings, sends it to Claude or to a model the organisation supplies, and applies the returned changes on a new branch as a draft pull request. Fixes are never merged automatically.

Repository access is a GitHub App with short-lived installation tokens; the model is not given the token. After a fix PR merges, a verification loop compares the issue against its subsequent event counts and resolves it once it stays silent for a configured window.

Open Dependabot findings on a connected repository can be turned into dependency-bump draft PRs through the same engine.

Each run is metered against the organisation's monthly allowance and a spend ceiling, with per-run cost derived from recorded model token usage.

Compatibility

Ingestion accepts the Sentry envelope and store formats, so an existing Sentry SDK reports to Condux by changing its DSN. OTLP/HTTP logs are also accepted, which lets an OpenTelemetry-instrumented application report without an SDK.

First-party SDKs are published for JavaScript/Node, browser, edge, Next.js, Python, Go, Ruby, PHP, JVM and .NET, with framework adapters for React, Express, Flask/Django/FastAPI, Rails, ASP.NET Core and Spring Boot. The SDKs are Apache-2.0; the server is FSL.

MCP endpoint

A read-only Model Context Protocol endpoint lets an AI agent debug against real production errors. Point Claude, Cursor or any MCP client at a project with a scoped token and it can list issues, read stack traces and page through occurrences. There are no write tools, and a token resolves to exactly one project.

Architecture

.NET 10 services and a Next.js / React / TypeScript dashboard.

SDKs ──envelope──▶ [Relay] ──▶ [Redpanda] ──▶ [Consumer] ──▶ ClickHouse (events, time series)
                                                          └──▶ PostgreSQL (issues, orgs, audit)
 Dashboard ──▶ [ControlPlane] ──▶ [Conductor] ──▶ draft PR

Valkey backs rate limiting and monthly quotas. The relay holds no state. Raw events expire per plan through a column-driven ClickHouse TTL rather than a scheduled job.

Repository layout

Shortened. The full README is on GitHub.