老師Roushi

Get started

Install Roushi

Four ways in. Same brain, same rules, same knowledge — pick the one that fits how you work.

01

Deploy to Vercel

Quickstart

Your own Roushi instance in under 5 minutes — free

One click deploys your own Roushi instance with a Neon Postgres database auto-provisioned. You own the data, you own the instance — completely isolated from everyone else.

Deploy with Vercel

1Click the button above

Vercel clones the repo into your GitHub account, provisions a Neon Postgres database, and prompts you for four keys:

  • OPENAI_API_KEY — for embeddings. Get one here (free tier works).
  • AUTH_SECRET — secures your sessions. Run openssl rand -base64 32 in your terminal.
  • ROUSHI_MCP_HTTP_TOKEN — protects your MCP API endpoint. Run openssl rand -hex 32.
  • CRON_SECRET — authenticates the automated maintenance jobs. Run openssl rand -hex 32.

2Visit your instance

Wait for the first build to complete (~60 seconds — Vercel auto-provisions Neon Postgres and runs db:setup as part of the build, so tables are created automatically). Then Vercel gives you a URL like roushi-yourname.vercel.app. Open it — the Ask interface is ready. No manual database setup required.

3Optional: clone locally for CLI access

Want to use the pnpm roushi CLI to ingest projects from your local machine, or work on your own customizations? Clone your fork:

git clone https://github.com/YOUR_USER/roushi-template
cd roushi-template
pnpm install
npx vercel link                  # link to your Vercel project
npx vercel env pull .env.local   # pull DATABASE_URL + other env vars
pnpm roushi ingest ./content     # ingest the starter content
pnpm roushi add-project ~/projects/some-project   # ingest your own

Skip this if you only want to use the web UI — your deployed instance works without any local setup.

4Enable GitHub Actions for auto-updates

Your repo ships with a daily workflow that opens a PR when upstream Roushi has new commits. GitHub disables Actions on new repos by default for security — to enable:

  1. Open your repo on GitHub
  2. Click the Actions tab
  3. Click "I understand my workflows, go ahead and enable them"

Once enabled, the "Sync from upstream" workflow runs daily at 12:00 UTC. When it finds new commits, you'll get a PR titled "Sync from upstream (roushi-template)" — review the diff, click Merge, Vercel redeploys.

Why not GitHub's "Sync fork" button? Vercel's deploy flow creates a brand-new independent repository in your account (not a true fork), so GitHub's UI sync button doesn't appear. The workflow above does the same job mechanically — it's the recommended path.

Don't want auto-updates? Skip this step. You can also pull updates manually any time using this GitHub compare URL github.com/YOUR_USER/YOUR_REPO/compare/main...samwsimpson:roushi-template:main — opens a PR with the diff for review.

5Optional: connect your editor

Your instance has an MCP endpoint at https://your-roushi.vercel.app/api/mcp. Set ROUSHI_MCP_HTTP_TOKEN in Vercel env vars, then follow the Claude Code or VS Code instructions below to connect.

Deployed? Head to /onboarding for the five things to do in your first ten minutes that make Roushi click.

02

Claude Code (MCP)

Deepest

Full integration — hooks, rules, proactive context

Claude Code gets the deepest integration: MCP tools, PreToolUse hooks that inject brain context before every edit, and automatic rule syncing into your workspace memory. This is the setup we use internally.

1Register the MCP server

Choose local (fastest) or remote (works from any machine):

# Local stdio — fast, no network
claude mcp add -s user roushi pnpm -- \
  --silent --dir /path/to/Roushi mcp

# Remote HTTPS — works anywhere
claude mcp add -s user roushi-remote \
  https://roushi.ai/api/mcp \
  --transport http \
  --header "Authorization: Bearer $YOUR_TOKEN"

2Verify

# In any Claude Code session:
mcp__roushi__roushi_search "your query here"

You should see brain results. All 35 MCP tools are now available in every session.

3Optional: enable the proactive hooks

The hooks make Roushi inject relevant context before you write code, not just when you ask. Contact us during the beta for the hook setup guide.

03

VS Code Extension

Works with Copilot, Cursor, Windsurf — any AI in VS Code

The VS Code extension talks to the same Roushi brain via the MCP HTTP endpoint. It syncs rules into your AI tool's instruction file and surfaces relevant context in a sidebar panel.

1Install the extension

code --install-extension kumokodo.roushi

Or search "Roushi" in the VS Code Marketplace.

2Configure your endpoint + token

Open VS Code Settings (Ctrl+,) → search roushi:

  • Endpoint: your instance URL — e.g. https://roushi-yourname.vercel.app/api/mcp
  • Token: the ROUSHI_MCP_HTTP_TOKEN you generated during deploy

3Open a workspace

The extension automatically:

  • Detects which product this workspace belongs to
  • Syncs applicable rules into .github/copilot-instructions.md (Copilot), .cursorrules (Cursor), and .windsurfrules (Windsurf)
  • Shows relevant brain context in the Advisor sidebar when you switch files

4Use the commands

Open the command palette (Ctrl+Shift+P):

  • Roushi: Ask the Master — synthesized answer with citations
  • Roushi: Search the Brain — hybrid search
  • Roushi: Sync Rules — manual rule re-sync
  • Roushi: Show Context — file-specific brain context
04

CLI only

Terminal-first — works with any editor, any AI, any workflow

The CLI is a standalone tool. No editor integration needed — just a terminal. Useful for scripting, CI, or when you want the brain without any IDE coupling.

1Clone and install

git clone https://github.com/samwsimpson/roushi-template
cd roushi-template
pnpm install
cp .env.example .env.local  # fill in DATABASE_URL, OPENAI_API_KEY
pnpm db:setup

2Use it

pnpm roushi search "auth patterns"
pnpm roushi think "how do we handle Drizzle ANY arrays?"
pnpm roushi graph-query roushi --hops 2
pnpm roushi rules list
pnpm roushi skill list

The Deploy to Vercel path is fully self-service — no token from us needed. You generate your own keys and own your instance.

Want a managed hosted instance instead? We're working on a SaaS tier — request early access and we'll reach out when it's ready.