Mushi Mushi

Mushi Mushi

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

Plain-English diagnosis + a paste-ready fix, right inside Cursor and Claude Code. No log-reading. No second LLM API key for MCP.

Fastest path — drop Mushi into your AI editor:

npx mushi-mushi setup --ide cursor   # or: --ide claude

Already shipping an app? One command installs the SDK + env vars + an optional test report:

npx mushi-mushi

Open source, self-hostable, MIT JS core — bring your own LLM key, no second key for MCP, no lock-in. Self-host in minutes · licensing.

What is Mushi, exactly? Read the one-page constitution: VISION.md — the single source of truth for positioning, the north-star sentence, and who this is for.

Node ≥22 · CI · SDK MIT · enterprise · Smithery

Vision · Quick start · Connect your editor · Self-host · Why not just Sentry? · Packages · Docs · Live demo · Operators / platform · Roadmap

↑ the diagnosis: plain-English root cause + a paste-ready fix prompt · click to open the live demo

---

60-second proof

Using Cursor, Claude Code, or another AI editor? Paste this prompt and your agent does the whole setup — install, wiring, and a verified test report:

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.

Or run the wizard yourself:

npx mushi-mushi

The wizard auto-detects your framework, installs the right SDK, writes framework-prefixed env vars (e.g. VITE_MUSHI_PROJECT_ID / VITE_MUSHI_API_KEY, or NEXT_PUBLIC_MUSHI_*) to .env.local, and prints the snippet to paste. Those two vars are all the SDK needs — no Supabase, no LLM key (see examples/sdk.env.example; the root .env.example is for self-hosting the backend only). Then, the moment something breaks:

1. The bug lands in your queue — screenshot, the user's note, the route, the last console + network events, device context.
2. Mushi produces the diagnosis: a plain-English root cause + a fix you can paste.
3. You pull it into your editor over MCP:

npx mushi-mushi setup --ide cursor    # then ask Cursor: "what's broken in prod?"

No Sentry, no account, no monitoring stack required to see value. Self-host the whole thing in under five minutes, or use the free hosted tier (no card).

---

What this is

For the solo AI-first builder (the vibe coder): you ship fast with Cursor, Claude Code, Lovable, or Bolt, then lose afternoons when something breaks in code you didn't fully write. Mushi is the comprehension layer — plain-English diagnosis in your editor over MCP, so a bug costs five minutes instead of your afternoon. (Small teams and agencies hit the same pain.)

These are the bugs your monitoring can't see, and the ones you didn't write:

• A user added a coupon and the pay button slipped under their keyboard.
• A new signup tapped Save twice because nothing visibly happened the first time.
• A Pro customer's dashboard takes 12 seconds to load — and they've opened the competitor's tab.
• A layout that looks fine on your laptop folds in half on the one Android model used by 18% of your traffic.

What it is not

Not another dashboard you have to go read, and not an enterprise monitoring stack — standalone first, no Sentry/Datadog/Firebase required. Full positioning: VISION.md.

---

The diagnosis loop

When a user shakes their phone (or clicks the reporter):

1. Capture — screenshot, route, user note, recent console + network events, device context.
2. Classify — two-stage LLM tags severity, category, and a plain-English root-cause hint. The screenshot goes through an air-gapped vision pass that can't see the text prompt. A nightly judge scores the classifier and feeds a prompt-A/B loop.
3. Connect — the report embeds into a knowledge graph (Postgres + pgvector). The same broken button reported twenty times shows up as one row, not twenty.
4. Fix (optional) — Dispatch fix (or Slack / MCP / CI) runs an agent in a sandbox, runs your tests, and opens a draft PR. You review it like any other PR.

flowchart LR
    subgraph App["Your app"]
        SDK["mushi-mushi/{react, vue, svelte, angular, …}<br/>shadow-DOM widget · screenshot · console · network"]
    end
    subgraph Edge["Supabase Edge (Hono gateway + ~50 functions)"]
        API["api"]
        FF["fast-filter"]
        CR["classify-report<br/>+ vision + RAG"]
        ORCH["fix-worker"]
    end
    subgraph DB["Postgres + pgvector"]
        REP["reports"]
        KG["knowledge graph"]
        FIX["fix_attempts"]
    end
    subgraph Agents["mushi-mushi/agents"]
        SBX["sandbox: e2b / modal / cloudflare"]
        GH["GitHub PR"]
    end
    SDK -->|HTTPS| API
    API --> FF --> CR
    CR --> KG
    CR --> REP
    REP --> ORCH --> Agents
    Agents --> GH

The architecture, sequence diagram, and component-by-component spec live in apps/docs/content/concepts/architecture.mdx.

---

Self-host in under 5 minutes

A single Docker Compose file gets you a working stack against your own Supabase project:

cd deploy
cp .env.example .env   # ANTHROPIC_API_KEY, Supabase creds
docker compose up -d

SELF_HOSTED.md and the Self-host in minutes guide are the long-form walkthroughs. A Helm chart lives at deploy/helm/ — one helm install on any cluster.

Hosted: sign up at kensaur.us/mushi-mushi/, click Start free, no card, create a project, and copy your projectId + apiKey. The free tier covers 50 diagnoses a month (no card required).

One BYOK rule, both ways. Self-host and you bring your own Anthropic / OpenAI key — you pay the vendor at list rate, we never mark up a token. On hosted you bring no key at all: we meter by diagnosis (the plain-English root cause + fix), never by tokens, with a per-project spend cap and 50 / 80 / 100% alerts so the bill can't surprise you. Full numbers: pricing.

Internal edge functions (fast-filter, classify-report, fix-worker, judge-batch, intelligence-report, usage-aggregator, generate-synthetic) authenticate via requireServiceRoleAuth. Never expose them with --no-verify-jwt. Only the public api function should face the internet — see packages/server/README.md.

---

Why not just Sentry?

The question every team asks, answered once:

Sentry tells you what threw. Mushi ingests that — plus the bugs that never throw — explains each one in plain English, and closes the loop with a fix your agent can ship. One queue, one audit trail, with or without Sentry.

Mushi works standalone. If you already run Sentry, point a Sentry issue-alert webhook at /v1/webhooks/sentry?projectId=<your-project> and errors land in the same queue as user reports — deduped per Sentry issue, triaged in plain English, fix-dispatchable from the console, Slack, or your editor. The loop closes both ways: merging a Mushi fix resolves the linked Sentry issue, and resolving in Sentry resolves the linked report. A repeat alert on a fixed report reopens it as a regression instead of filing a duplicate.

Beyond Sentry, inbound adapters translate Datadog, Bugsnag, Rollbar, Crashlytics, New Relic, Honeycomb, Grafana Loki, CloudWatch, Opsgenie, and Firebase alerts into the same queue; outbound plugins keep Linear, Jira, GitHub Issues, Slack, Discord, Teams, and PagerDuty in sync. Nothing gets ripped out — Mushi is the mediator between what you already run. Operator-depth (SSO, retention, region routing): docs/operators/.

	Mushi	Sentry	Langfuse
Catches	Thrown errors and silent UX bugs (dead clicks, slow screens, layout breaks)	Thrown errors, performance traces	LLM call traces, prompt evals
Output	Plain-English root cause + paste-ready fix, in your editor	Stack trace + breadcrumbs, in a dashboard	Trace tree + scores, in a dashboard
Auto-fix	Optional: sandbox agent opens a draft PR	Seer add-on (paid)	Not in scope
Second LLM key for MCP	No — reuses your app's key	N/A	N/A
Setup	One command, no account required to try	SDK + DSN + dashboard	SDK + project + dashboard

Different jobs: Sentry watches what your code throws, Langfuse watches what your LLM calls do, Mushi watches what your user experiences — including the bugs that never throw.

---

Framework coverage

Most developers install one SDK — npx mushi-mushi picks it for you. React/Next.js quick start:

npm install @mushi-mushi/react      # also covers Next.js

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

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

// Vue 3 / Nuxt
import { MushiPlugin } from '@mushi-mushi/vue';
app.use(MushiPlugin, { projectId: 'proj_xxx', apiKey: 'mushi_xxx' });

// Svelte / SvelteKit
import { initMushi } from '@mushi-mushi/svelte';
initMushi({ projectId: 'proj_xxx', apiKey: 'mushi_xxx' });

// Angular 17+
import { provideMushi } from '@mushi-mushi/angular';
bootstrapApplication(AppComponent, { providers: [provideMushi({ projectId: 'proj_xxx', apiKey: 'mushi_xxx' })] });

// React Native / Expo
import { MushiProvider } from '@mushi-mushi/react-native';

// Vanilla JS / any framework
import { Mushi } from '@mushi-mushi/web';
Mushi.init({ projectId: 'proj_xxx', apiKey: 'mushi_xxx' });

iOS (Swift PM, v0.4.0): .package(url: "https://github.com/kensaurus/mushi-mushi.git", from: "0.4.0") · Android (Gradle): dev.mushimushi:mushi-android:0.4.0 · Flutter: pub add mushi_mushi.

Want a runnable example? examples/react-demo is a minimal Vite + React app with test buttons for dead clicks, thrown errors, failed API calls, and console errors.

Full package list and maturity table: SDK reference.

---

Where it stops

Mushi is honest about what's still partial. Skim before you commit:

Area	Working	Still partial
Classification	Haiku fast-filter, Sonnet deep + vision air-gap closed, structured outputs, prompt-cached prompts, pg_cron self-healing, Stage 2 streaming via streamObject with progressive reports.stage2_partial UI updates and OpenAI fallback	—
Judge / self-improve	Sonnet judge with OpenAI fallback, prompt A/B auto-promotion via judge → avg_judge_score → promoteCandidate, OpenAI fine-tune adapter end-to-end (submit JSONL → poll → predict against fine_tuned_model_id, BYOK OPENAI_API_KEY), Bedrock fine-tune adapter (SigV4-signed CreateModelCustomizationJob, requires MUSHI_BEDROCK_FINETUNE_ENABLED=1 + AWS BYOK keys)	Anthropic fine-tune API is not publicly self-service in 2026 — the adapter stub links to the access-request form.
Fix orchestrator	Single-repo validateResult gating, GitHub PR, MCP JSON-RPC 2.0 client, multi-repo coordinator, first-party ClaudeCodeAgent (spawns local claude CLI) and CodexAgent (OpenAI Responses API, BYOK) — both gated behind explicit env flags so shared deployments never invoke them unintentionally	—
Sandbox	Provider abstraction; local-noop (tests) + e2b / modal / cloudflare (prod). Production refuses local-noop unless MUSHI_ALLOW_LOCAL_SANDBOX=1.	—
Verify	Playwright screenshot diff + step interpreter (navigate / click / type / press / select / assertText / waitFor / observe)	—
Enterprise	Plugin marketplace + HMAC, audit ingest, region pinning, retention CRUD, Stripe metering, SAML SSO via Supabase Auth Admin API, OIDC SSO self-service — see the commercial boundary below for which of these are paid/Enterprise-tier	—
Graph backend	SQL adjacency over graph_nodes / graph_edges ships in every deployment	Apache AGE is a hosted-tier enhancement when the extension is installed. Managed Supabase stays on SQL adjacency.
Inventory v2 & QA-gates	Hand-written inventory.yaml, SDK-driven discovery, Claude proposer, ESLint gate rules, 5-gate composite GitHub check, synthetic monitor, expected_outcome contract end-to-end — see docs/operators/	Inventory is gated behind Advanced mode + the inventory_v2 plan flag.
Self-host (Helm)	Single-pod deploy on any Kubernetes; pre-install Job applies all SQL migrations from a bundled ConfigMap. Multi-region via global.region + global.peerRegions Helm values.	Full active/active write replication is not automated yet — write routing relies on client-side region stickiness.

---

Running this for a team?

The platform depth — inbound adapters, outbound plugins, A2A / AG-UI / MCP interop, the inventory.yaml QA-gate system, the synthetic monitor, SSO / audit / retention / region pinning, and open-standards plumbing — lives in docs/operators/ so the front door stays on the wedge. Start there if you're wiring Mushi into an existing stack or evaluating it as a platform.

---

Cursor & Claude Skills

Install Mushi skills in your Cursor or Claude Code project for one-command setup, usage, and debugging:

npx skills add kensaurus/mushi-mushi

Then: /mushi-setup (guided SDK install + MCP wiring), /mushi-debug (diagnose ingest / MCP / pipeline failures), /mushi-health (pass/fail check across CLI, API, edge functions, BYOK keys), /mushi-integration (two-way loop, fix dispatch, lessons). The admin Connect & Update page (/connect) mirrors the same flows with one-click Add to Cursor deeplinks.

Repo at a glance (run pnpm docs-stats): ~373K TS lines · 1,772 source files · 44 workspace / 36 npm packages · 55 edge functions · 337 SQL migrations · 19 pipeline agents. Full tour: docs/SCREENSHOTS.md.

---

Contributing

Issues and PRs welcome:

git clone https://github.com/kensaurus/mushi-mushi.git
cd mushi-mushi
pnpm install
pnpm dev

Requires Node.js ≥ 22 and pnpm ≥ 10. See individual package READMEs, docs/stats.md for canonical counts, and CONTRIBUTING.md.

License & branding

This repository is open-core — the Supabase / Grafana model. The SDK packages are MIT — use them in any product, open or closed. The server (the part you self-host or we run for you) is AGPLv3 — true OSI open source: self-host it, fork it, modify it for your own org. If you offer a modified server as a hosted service to third parties, publish your changes or see COMMERCIAL-LICENSE.md. A small Enterprise Edition boundary (packages/server/ee/) is source-available but commercial for production use — that's operator/enterprise plumbing only, never the wedge.

Surface	License	Permitted	Notes
SDK packages — core, web, react, vue, svelte, angular, react-native, capacitor, flutter, ios, android, node, cli, mcp, mcp-ci, plugin-* (13 plugins), adapters (11 sources), inventory-schema, inventory-auth-runner, eslint-plugin-mushi-mushi, brand, marketing-ui	MIT	Use, fork, sell, embed in proprietary products.	Trademarks separate — see below.
Server packages — @mushi-mushi/server, @mushi-mushi/agents, @mushi-mushi/verify	AGPLv3	Use, modify, self-host, fork for your own org. SaaS modifiers publish changes or commercial license.	OSI-approved copyleft. The cloud runs this exact core.
Enterprise features — SSO/SCIM, audit-log ingest, retention policy CRUD, region pinning, SOC2 evidence	Commercial / paid tier	Available on the Enterprise plan (hosted) or with a commercial license (self-host).	The code may be source-visible, but production use of these specific features is a paid boundary — see docs/operators/.
Trademarks — "Mushi Mushi", "Mushi", 虫, the bug logo	Trademark policy	Refer to the project, build add-ons, link to the repo.	Forks must rename. Hosting a service under the Mushi name requires written permission.
Third-party attributions	NOTICE	—	Upstream projects we depend on and their licenses.

Security researchers: see SECURITY.md for the threat model, PII commitments, and safe-harbor terms.

---

Also by @kensaurus

Other free apps and tools from the same Tokyo studio:

App	What it does	Links
glot.it — Learn Thai Free	161 lessons, pitch-contour tone mirror, AI roleplay chat, offline-first.	App Store · Google Play
yen-yen — Expense Tracker	Kakeibo-style household ledger. No bank password, no ads, no auto-writes.	App Store · Google Play
The Wanting Mind — Free Book	147,000-word interactive book — 3D knowledge graph, 12 narrators, 22 simulations.	App Store · Google Play
cursor-kenji	58 Cursor AI agent skills for React / Next.js / Supabase development.	npx skills add kensaurus/cursor-kenji

---