Mushi Mushi

Mushi Mushi

Your AI wrote it. Mushi tells you why it broke.

When a user hits a bug, Mushi explains the cause in plain English and hands your editor a fix to start from.

Built for the vibe coder who ships fast with Cursor, Claude Code, Lovable or Bolt, so a bug costs five minutes instead of an afternoon.

Quick start · How it works · Docs · Demo · Why not Sentry? · Self-host · Vision

↑ a diagnosis: the root cause in plain English and a paste-ready fix prompt. Click to open the read-only demo.

How it works

flowchart LR
    U["Your users<br/>tap the bug widget"] --> M
    S["Your monitoring<br/>Sentry, Crashlytics… (optional)"] -.-> M
    M["Mushi<br/>one queue · diagnosis · fix"] --> E["Your editor or agent<br/>Cursor, Claude Code (MCP)"]
    E --> P["Draft PR<br/>you review it"]
    P --> R["Release<br/>report closes, user is told"]

1. Capture. A user taps the widget. Mushi records a screenshot, the route, their note, and the last console and network events.
2. Diagnose. Mushi writes the root cause in plain English with a fix prompt. Twenty reports of the same broken button become one row.
3. Fix. Your editor reads the diagnosis over MCP. If you want, an agent opens a draft PR that you review like any other. A second model (Claude Opus, on your own key) reviews each agent PR first, and Mushi will not merge one it says to hold.

It catches the bugs that never throw an error: the pay button hidden under the keyboard, the Save button tapped twice because nothing happened, the dashboard that takes 12 seconds.

Quick start

npx mushi-mushi                      # detects your framework, installs the SDK, writes env vars, sends a test report
npx mushi-mushi setup --ide cursor   # or --ide claude: your editor can now read diagnoses

Then ask your editor: "what's broken in prod?"

The SDK needs two env vars (VITE_MUSHI_PROJECT_ID / VITE_MUSHI_API_KEY, or NEXT_PUBLIC_MUSHI_*; see examples/sdk.env.example). No Sentry, no LLM key, no monitoring stack. The hosted free tier covers 50 diagnoses a month with no card, or you can self-host.

Let your AI editor do the setup

Paste this into Cursor or Claude Code:

Install the Mushi skills from github.com/kensaurus/mushi-mushi (npx skills add kensaurus/mushi-mushi), then run the mushi-setup skill to wire the Mushi SDK and MCP server into this app and send a test report to verify the connection.

The skills also add /mushi-debug (why ingest or MCP fails), /mushi-health (pass/fail across CLI, API and keys) and /mushi-integration (fix dispatch and the two-way loop).

Add it by hand (React shown; Vue, Svelte, Angular, React Native, vanilla JS, native)

import { MushiProvider } from '@mushi-mushi/react';

export function App() {
  return (
    <MushiProvider config={{ projectId: 'proj_xxx', apiKey: 'mushi_xxx' }}>
      <YourApp />
    </MushiProvider>
  );
}

import { MushiPlugin } from '@mushi-mushi/vue';        // Vue 3 / Nuxt: app.use(MushiPlugin, { projectId, apiKey })
import { initMushi } from '@mushi-mushi/svelte';       // Svelte / SvelteKit: initMushi({ projectId, apiKey })
import { provideMushi } from '@mushi-mushi/angular';   // Angular 17+: providers: [provideMushi({ projectId, apiKey })]
import { MushiProvider } from '@mushi-mushi/react-native';
import { Mushi } from '@mushi-mushi/web';              // any framework: Mushi.init({ projectId, apiKey })

Native SDKs (iOS, Android, Flutter) are a preview and install from this repository: see packages/ios, packages/android and packages/flutter. A runnable demo lives in examples/react-demo.

What's inside

Full list and maturity: SDK reference. Architecture: how the pipeline fits together.

Agent PRs get an independent review before anyone merges them: merge, fix first or hold, with findings on the PR. A "fix first" PR can go back to the agent once with those findings (agent PR review).

Fix waiting to land? mushi runner status shows where your tests run, why they wait, and whether a red check was your code or a test machine that stopped (CI runners).

Why not just Sentry?

Sentry is built around what the code threw, with a User Feedback widget and replay alongside. Mushi starts from what the user reported, ingests Sentry's errors too, explains each one in plain English, and hands your agent a fix prompt to start from. One queue, with or without Sentry.

It works with what you already run. Point a Sentry alert webhook at /v1/webhooks/sentry?projectId=<id> and errors land in the same queue. Merging a Mushi fix resolves the Sentry issue, and the reverse. Adapters also bring in Datadog, Bugsnag, Rollbar, Crashlytics and others. Plugins keep Linear, Jira, GitHub Issues, Slack, Discord, Teams and PagerDuty in sync. Nothing gets ripped out. Details: Sentry vs Mushi.

Self-host

cd deploy && cp .env.example .env   # add ANTHROPIC_API_KEY and your Supabase credentials
docker compose up -d

Walkthroughs: SELF_HOSTED.md and Docker Compose guide. Self-hosting is bring-your-own-key: you pay your LLM vendor at list price. The hosted service needs no key; it bills per diagnosis with a spend cap (pricing). Teams wiring Mushi into an existing stack (SSO, retention, region routing, Kubernetes) start at docs/operators/. Only the public api function should face the internet: see internal caller authentication.

Where it stops

Repo at a glance (run pnpm docs-stats): ~570K TS lines · ~2,500 source files · 45 workspace / 37 npm packages · 76 edge functions · 506 SQL migrations · 19 pipeline agents. Full tour: docs/SCREENSHOTS.md.

Community and contributing

Questions and show-and-tell: GitHub Discussions. Bugs: Issues (first reply within 24 hours on weekdays). Security: SECURITY.md. Anything else private: kensaurus@gmail.com.

git clone https://github.com/kensaurus/mushi-mushi.git && cd mushi-mushi
pnpm install && pnpm dev   # Node ≥ 22, pnpm ≥ 10

See CONTRIBUTING.md and docs/stats.md.

License

Open core. Each package directory's own LICENSE governs that package; the root LICENSE states the split.

Self-hosting is free and complete: the repo has every feature and working default prompts. Mushi Cloud runs the same code and keeps its own tuning data (prompt wording, evaluation sets and scoring weights tuned on live traffic) as configuration. Contributions come in under a CLA. Why it is set up this way: ADR 0036. Third-party notices: NOTICE.

More apps from the same maker

All apps live under kensaur.us.

If Mushi helped, star the repo so the next vibe coder finds it. Open an issue · Follow on Bluesky