DebugBundle

Production debugging bundles for AI agents, with runtime error reporting and incident response.

DebugBundle provides runtime error reporting, crash reporting, incident response, endpoint health checks, and product analytics for humans and AI agents. It captures runtime failures, groups them into incidents, and publishes deterministic debug bundles. Its monitoring scope is customer-facing runtime behavior and endpoint health, not generic infrastructure metrics.

Works with AI coding agents including Codex, Claude Code, and Gemini CLI, as well as Cursor and GitHub Copilot. See the dedicated Codex, Claude Code, and Gemini CLI setup guides for native packages and direct MCP connections.

Why DebugBundle?

Modern AI agents are useful only when they get enough trustworthy context. DebugBundle packages the facts around a production incident into a versioned bundle instead of leaving agents to scrape dashboards, logs, traces, and chat threads.

Key properties:

• Agent-native bundles: deterministic failure and improvement bundles with errors, requests, responses, logs, frontend context, deploy metadata, runtime details, and reproduction hints.

• Interface parity: API, CLI, and MCP expose the same incident, bundle, probe, webhook, alert, project, and automation workflows.

• Local-first setup: start without a cloud account by writing events to .debugbundle/local/events/, then connect to DebugBundle Cloud when ready.

• Safe SDKs: SDK failures are swallowed internally, sensitive fields are redacted before transport, and duplicate storms are suppressed locally.

• Self-hostable core: Compose-based stack for the web app, API, worker, Postgres, Redis, and S3-compatible object storage.

Related MCP server: AIOps MCP

AnalyticsBundle

AnalyticsBundle extends debugging from incident evidence to product-usage evidence without turning DebugBundle into a long-term raw-event store. It is opt-in browser analytics for the questions a human or agent needs to improve a product: visits and active users, routes and funnels, device/browser/OS/language segments, feature use, friction markers, incident impact, and bounded structured journey replay.

• Ask directly: API, CLI, and MCP expose aggregate metrics, journey patterns, opportunities, and generated AnalyticsBundles through the same project-authorized surface.

• Generate by analysis unit: bundles describe a usage, funnel, route, friction, conversion, deploy, or incident-impact question. DebugBundle does not create one AnalyticsBundle per visit.

• Keep evidence explainable: a generated bundle includes aggregate metrics, linked incidents/deploys, privacy-safe journey timelines, and deterministic journey-selection rank/basis for agent review.

• Stay privacy- and cost-conscious: raw analytics inputs and retained journey samples expire; long-lived usage is aggregate rollups. Debug capture remains independent when analytics is disabled, unavailable, sampled out, quota-limited, or unhealthy.

See the repository public interface contract for API/CLI/MCP parity and the self-host guide for retention and upgrade behavior.

Quick Start

Choose the path that matches how you want to evaluate DebugBundle.

Cloud

Use Cloud when you are preparing a hosted deployment or want team-visible incidents, alerts, webhooks, GitHub automation, API access, and MCP access.

npm install -g @debugbundle/cli
debugbundle setup
debugbundle login
debugbundle connect

debugbundle connect creates or selects a cloud project, creates a write-only project token, and updates .debugbundle/local/connection.json. Put the shown project token in your hosted environment:

DEBUGBUNDLE_PROJECT_TOKEN=dbundle_proj_xxxxxxxxxxxx

Add the smallest SDK or ingestion path that matches your app, deploy it with the token configured, then verify ingestion:

debugbundle verify cloud --project-id proj_01HXYZ... --trigger-5xx
debugbundle incidents --source cloud
debugbundle inspect inc_01HXYZ...

See the full Cloud quickstart and connect-to-cloud guide.

Local-only

Use local-only mode when you want captured data and bundles to stay on the machine or storage volume where the SDK and CLI run.

npm install -g @debugbundle/cli
debugbundle setup --project-mode local-only

Initialize an SDK in local mode where supported, or use debugbundle watch for existing logs. After triggering a test error:

debugbundle process
debugbundle incidents --source local
debugbundle inspect inc_local_...

Local events are written under .debugbundle/local/events/; generated bundles are written under .debugbundle/bundles/. See the local-only guide.

Install an SDK

All SDKs follow the same universal interface: init, captureException, captureError, captureLog, captureRequest, captureMessage, setContext, probe, and flush.

Node.js

npm install @debugbundle/sdk-node

import { debugbundle } from "@debugbundle/sdk-node";

debugbundle.init({
  projectToken: process.env.DEBUGBUNDLE_PROJECT_TOKEN,
  environment: "production",
  service: "api"
});

debugbundle.captureExceptions();
debugbundle.captureRejections();

Express, Fastify, Next.js, pino, winston, bunyan, local file transport, remote capture policy, probes, and browser relay handlers are supported.

Browser

npm install @debugbundle/sdk-browser

import { createDebugBundleBrowserSdk } from "@debugbundle/sdk-browser";

const debugbundle = createDebugBundleBrowserSdk();

debugbundle.init({
  transportMode: "relay",
  endpoint: "/debugbundle/browser",
  environment: "production",
  service: "web"
});

For full-stack apps, prefer a backend browser relay so project tokens stay server-side. Same-origin relay paths are simplest; split frontend/backend deployments can use explicit browser relay mode with an absolute backend relay URL and backend origin allowlisting. Frontend-only deployments can send directly to DebugBundle Cloud with a dedicated public write-only token and an allowed-origin restriction. See Browser Relay Setup.

Python

pip install debugbundle-python

import os
import debugbundle

debugbundle.init(
    project_token=os.environ["DEBUGBUNDLE_PROJECT_TOKEN"],
    environment="production",
    service="api",
)

debugbundle.capture_exceptions()
debugbundle.capture_logging()

Django, Flask, FastAPI, Python logging, structlog, loguru, local file transport, remote capture policy, probes, and browser relay helpers are supported.

PHP

composer require debugbundle/sdk-php

<?php

use DebugBundle\DebugBundle;

DebugBundle::init([
    'projectToken' => getenv('DEBUGBUNDLE_PROJECT_TOKEN'),
    'environment' => 'production',
    'service' => 'api',
]);

DebugBundle::captureErrors();
DebugBundle::captureExceptions();
DebugBundle::captureShutdown();

Laravel, Symfony, Monolog, local file transport, remote capture policy, probes, and browser relay adapters are supported.

Ruby

gem install debugbundle

require "debugbundle"

DebugBundle.init(
  project_token: ENV["DEBUGBUNDLE_PROJECT_TOKEN"],
  environment: "production",
  service: "api"
)

DebugBundle.capture_exceptions

Rails, Rack, Sidekiq, Ruby Logger, Semantic Logger, local file transport, remote capture policy, probes, and browser relay handlers are supported.

Java

<dependency>
  <groupId>com.debugbundle</groupId>
  <artifactId>debugbundle-spring-boot-starter</artifactId>
  <version>0.1.0</version>
</dependency>

debugbundle:
  project-token: ${DEBUGBUNDLE_PROJECT_TOKEN}
  environment: production
  service: api
  project-mode: connected

The Spring Boot starter supports servlet request capture, MVC exception capture, Logback capture, remote config, probes, and an optional browser relay route.

Go

go get github.com/debugbundle/debugbundle-go

client := debugbundle.New(debugbundle.Config{
  ProjectToken: os.Getenv("DEBUGBUNDLE_PROJECT_TOKEN"),
  Environment:  "production",
  Service:      "api",
})
defer func() { _ = client.Flush(context.Background()) }()

net/http, Gin, Echo, slog, zap, zerolog, local file transport, remote capture policy, probes, and browser relay handlers are supported.

WordPress

Install DebugBundle from the WordPress.org plugin directory, then open Settings -> DebugBundle and save your project token. The plugin bundles backend PHP capture, frontend browser capture, and a WordPress REST relay so the project token stays server-side.

CLI, API, and MCP

The CLI is the daily operational entry point:

npm install -g @debugbundle/cli
debugbundle setup
debugbundle doctor
debugbundle verify local
debugbundle verify cloud --trigger-5xx
debugbundle incidents
debugbundle inspect <incident-id>

Automation can use the HTTP API directly or the MCP server for agent workflows:

• API reference: https://debugbundle.com/docs/api

• CLI reference: https://debugbundle.com/docs/cli

• MCP docs: https://debugbundle.com/docs/mcp

• MCP distribution channels: https://debugbundle.com/docs/mcp/distribution

• OpenAI Plugin candidate: https://debugbundle.com/docs/mcp/openai-plugin

• Bundle schema: https://debugbundle.com/docs/bundles/schema

Marketplace-managed MCP clients can run npx @debugbundle/mcp and provide DEBUGBUNDLE_MEMBER_TOKEN in the MCP server environment. The official MCP Registry name is com.debugbundle/mcp; project tokens are SDK write-only ingestion credentials and must not be used for MCP retrieval or management.

The separate OpenAI Plugin 1.0.0 production candidate combines a tailored skill with an OAuth-protected twenty-three-tool read-only remote projection plus the owner-approved existing-app consent, synthetic-reviewer, and Settings revocation surfaces. Its nine analytics tools expose bounded aggregate usage, route, device, acquisition, action, funnel, journey-pattern, and incident-impact metrics while excluding individual journeys, custom dimensions, analytics bundles/opportunities, and mutations. It is active at the permanent https://mcp.debugbundle.com/mcp origin for owner-approved Developer Mode validation, preserves the stdio/OpenClaw surface, and is not submitted, published, or publicly installable.

For local visual review without a real provider interaction, run make dev-openai-plugin-preview and open http://localhost:5291/__dev/openai-plugin. The opt-in development route uses the production UI components with deterministic synthetic data and provides every consent/reviewer/Settings state, all 64 scope subsets, and 390 px, 768 px, and 1280 px iframe viewports. Its actions stay in browser memory and never call OAuth, reviewer, grant, or revocation APIs. The route is absent from production builds, and preview evidence does not replace manual accessibility, MCP Inspector, outside-network reviewer, ChatGPT Developer Mode, deployed, submission, or publication validation.

Repository Layout

apps/
  api/       Fastify ingestion and retrieval API
  worker/    BullMQ processing worker for normalization, grouping, bundles, alerts, and webhooks
  cli/       @debugbundle/cli command-line interface
  mcp/       @debugbundle/mcp server for agent workflows
  web/       React/Vite app for interactive project and incident management
packages/
  auth/              Auth, sessions, token generation, token hashing
  bundle-engine/     Deterministic bundle assembly
  event-normalizer/  Event validation, normalization, classification, fingerprinting
  log-parser/        CLI log ingestion parser registry
  redaction/         Sensitive data scrubbing
  retrieval-client/  Shared retrieval API client used by CLI and MCP
  shared-types/      Zod schemas, TypeScript types, bundle/event contracts
  storage/           Postgres, Redis, S3-compatible storage adapters and migrations
sdks/
  debugbundle-js/         Local clone of the JS SDK repo
  debugbundle-python/     Local clone of the Python SDK repo
  debugbundle-php/        Local clone of the PHP SDK repo
  debugbundle-java/       Local clone of the Java SDK repo
  debugbundle-go/         Local clone of the Go SDK repo
  debugbundle-wordpress/  Local clone of the WordPress plugin repo
  debugbundle-ruby/       Local clone of the Ruby SDK repo
site/
  Public docs, marketing, reference, and blog site clone

The SDKs are standalone repositories under the debugbundle GitHub organization. This core repository owns the product services, shared contracts, CLI/MCP surfaces, and core-owned shared JS packages.

Local Development

Use the Make targets so routine commands run in Docker-scoped environments.

make install
make infra-up
make infra-bootstrap
make dev

Local services:

Useful checks:

make lint
make typecheck
make test
make build
make ci

make dev requires DEBUGBUNDLE_PROBE_TRIGGER_SECRET and ANALYTICS_HASH_SECRET in .env. Start from .env.example, then keep local-only overrides in .env.local when needed.

Populated UI preview

After make install, run make dev-mock and open http://localhost:5291/dashboard. It restarts only the web container and automatically signs the preview in as demo@example.test. Synthetic projects, incidents, improvements, 30-day health history, alert/capture rules, webhook endpoints and delivery history, project and member tokens, members/invitations, billing, provider connections, and GitHub deliveries are served locally, so the data works in your ordinary browser. SayCheese has analytics enabled with sample funnels, flows, journeys, opportunities and generated artifacts. TaskTime includes empty and analytics-disabled states.

Project edits, incident/improvement actions, rule and health-check edits, webhook creation/tests, token creation/revocation, invitations/member edits, probe activation, weekly reports, repository selection, and analytics settings/flows/ funnels/bundle generation are simulated in memory. Failed-delivery clearing uses the normal browser-only behavior. Billing checkout and portal links stay in the local preview; capacity changes only update mock state. Mock credentials are deliberately unusable with real installations. No real checks, notifications, GitHub dispatches, OAuth interactions, payments or database writes are made. Unknown API operations still return an explicit preview error without contacting the backend. External provider authorization and private operator tools require normal development. After a mock logout, use any valid email and code 123456 to return to the demo account; the same code is used for simulated account deletion.

• Run make dev-mock again to reset the simulated server data. Browser preferences and cleared-delivery IDs remain in local storage; use Show cleared to inspect them.

• Run make dev-mock-off to return the frontend to your real local API.

• The preview binds to loopback, blocks foreign origins and intercepts unknown API paths rather than forwarding them. It requires development serve mode, is absent from production builds and disables frontend telemetry even if local env enables it.

Self-Hosting

The supported self-host bootstrap lives in deploy/selfhost/.

git clone https://github.com/debugbundle/debugbundle.git
cd debugbundle/deploy/selfhost
cp .env.example .env
docker compose up -d

The self-host stack includes the web app, API, worker, PostgreSQL, Redis, and LocalStack S3. See Self-Hosting and deploy/selfhost/README.md.

Documentation

• Public docs: https://debugbundle.com/docs

• Quickstart: https://debugbundle.com/docs/quickstart

• Installation: https://debugbundle.com/docs/installation

• SDKs: https://debugbundle.com/docs/sdks

• Agent workflows: https://debugbundle.com/docs/agent-workflows

• System overview: SYSTEM_OVERVIEW.md

• Architecture map: ARCHITECTURE_MAP.md

• Requirements: spec/requirements.md

• Acceptance criteria: spec/acceptance.md

• Public interfaces: contracts/public-interfaces.md

Release Model

The canonical public product release is the root debugbundle repository tag and GitHub Release (v*). Package-specific releases are separate:

• cli-v* for @debugbundle/cli

• mcp-v* for @debugbundle/mcp

• shared-js-v* for @debugbundle/shared-types and @debugbundle/redaction

Standalone SDK repositories publish and version their own release surfaces independently:

• debugbundle-js for @debugbundle/sdk-node and @debugbundle/sdk-browser

• debugbundle-python for debugbundle-python

• debugbundle-php for debugbundle/sdk-php

• debugbundle-java for Maven artifacts

• debugbundle-go for Go modules

• debugbundle-wordpress for the WordPress plugin

The v1 release train publishes dependency roots before dependent wrappers:

1. Publish @debugbundle/shared-types and @debugbundle/redaction from the core repo first.

2. Publish @debugbundle/sdk-node and @debugbundle/sdk-browser from debugbundle-js after the matching shared-package version exists on npm.

3. Publish independent SDK and package families whose artifacts do not bundle another DebugBundle SDK.

4. Publish React Native after the Android and iOS native SDK versions it delegates to are live and smoke-tested.

5. Publish WordPress after the PHP SDK and browser SDK versions it requires are live and smoke-tested, then rebuild the bundled browser asset.

6. Bump hosted dogfooding manifests only after the referenced registry versions exist.

7. Create the canonical core GitHub Release after package-specific release workflows pass.

Our own hosted/source-deployed dogfooding surfaces intentionally consume published packages rather than implicit workspace links. After a successful registry publish, bump the pinned versions in the root package.json, hosted app apps/web/package.json, and public-site site/package.json before running hosted validation or deployment.

Contributing

Read CONTRIBUTING.md before opening a pull request. The short version:

• Keep app/package boundaries strict.

• Add or update tests for behavior changes.

• Run make lint, make typecheck, make test, and make build before asking for review.

• Update docs, contracts, and public interface references when behavior changes.

Security

Do not report vulnerabilities in public issues. Use GitHub private vulnerability reporting for this repository:

https://github.com/debugbundle/debugbundle/security/advisories/new

See SECURITY.md for scope and response expectations.

License

DebugBundle core, SDKs, and executable developer tools use Apache-2.0. Service-required exceptions are limited to the WordPress plugin (GPL-2.0-or-later) and portable ClawHub/Smithery instruction skill (MIT-0). See the licensing policy.

Shared shadcn styles

The web app vendors the unchanged shadcn 4.1.2 Tailwind stylesheet in apps/web/src/styles/shadcn.css, including its upstream MIT license. This preserves the current component animations and data variants without installing the component generator and its unused vulnerable glob-parser dependency in the app. Existing shadcn components and components.json remain the source of the UI. Run the component generator separately through its package runner when needed, and review any generated dependency/import changes against the dependency audit before adoption.

make audit checks the installed dependency graph for high and critical advisories; core CI and core publication enforce the same threshold. Dependency refreshes use bounded vulnerable-version overrides and preserve independently published SDK pins.