Roushi · Private beta
"Ask the master." A self-wiring knowledge layer and autonomous coordination plane for portfolio operators. Roushi is the oracle, not an assistant — every decision, pattern, lesson, and goal across all your products lives here, and it prescribes context to Claude Code before you write code, not after. Built on Next.js 16, Neon Postgres + pgvector, Drizzle, Auth.js v5, and a custom MCP server (stdio + HTTPS).
Solo founders and portfolio operators carry the entire context of their business in their head. As the portfolio grows past five products, that context fragments — decisions in product A are forgotten when working on product B, lessons from a painful incident in repo X don't surface when the same trap appears in repo Y, and every AI agent (Claude Code, Cursor, Windsurf) cold-starts each session with no memory of prior work.
What's broken
What we built
Six layers, one purpose: the brain knows what you've built, and Claude Code uses that knowledge before generating each next file. Roushi is wired into every project workspace via PreToolUse hooks that inject relevant context the moment you start editing.
The homepage Ask flow runs your question through hybrid search (top-12 by default), grounds an LLM (gpt-4o) on the retrieved entities, and returns a synthesized answer with [slug] citations linking to entity detail pages. Below the answer: a gap analysis — what the LLM judged the brain doesn't know yet, which becomes leads for what to ingest next.
Before every Edit/Write in any project workspace, a hook classifies the file by path pattern, looks up the most relevant lessons/decisions/patterns/rules from the brain, and injects them as context. Claude reads the prior knowledge before generating the code. Result: code lands secure and idiomatic on the first try.
Twelve entity types (product, tech, vendor, decision, pattern, incident, lesson, goal, person, roadmap_item, rule, skill) connected by ten typed relations (uses, depends_on, hosted_on, defined_by, led_to, mentions, applies_to, blocks, shares_pattern, supersedes). The /graph page renders the full portfolio knowledge as a force-directed view; click any node to open its entity.
Four scans run on schedule: decay (daily 3am UTC) flags entities unvalidated >90d; audit-edges (Sun 4am) flags wrong relations; dedupe (1st of month) finds near-duplicate slugs; contradictions (Sun 6am) uses pgvector close-pair search + LLM judge to flag conflicting rules/decisions/lessons. Each scan posts findings as [system] goals; goals auto-close when findings clear.
Single Next.js 16 app at the repo root — web UI, API routes, Server Actions, MCP route, and CLI all live in one tree. Turborepo split deferred until CLI/SDK shape stabilizes.
Vercel Fluid Compute with Node 24 LTS for the API surface. Four Vercel Cron jobs invoke maintenance scans on schedule. The MCP server ships as both a stdio binary (pnpm mcp) and a deployed HTTPS Route Handler at /api/mcp.
Every brain capability is a typed MCP tool. Local stdio is registered globally in Claude Code for low-latency access; the deployed HTTPS endpoint exposes the same 35 tools to remote agents.
Neon Postgres with pgvector for embeddings (1536-dim text-embedding-3-small) and tsvector for full-text. Drizzle ORM with snapshot migrations. Hybrid search fuses both rankings via Reciprocal Rank Fusion (k=60).
LLM-synthesized answers grounded in the brain. Citations and gap analysis on every response.
pgvector cosine + Postgres tsvector keyword, fused via Reciprocal Rank Fusion. Optional LLM rerank stage.
Force-directed cytoscape view of every entity and edge. Type-filter legend, name search, click-to-navigate.
12 entity types and 10 typed relations. Multi-hop traversal via recursive CTE with cycle protection.
Roushi-as-advisor hook injects relevant brain context into Claude Code before every Edit/Write.
File-path-pattern security guidance fires before code lands. Composes with portfolio-security-baseline rule.
Rules in the brain auto-write into every workspace's Claude memory dir via the SessionStart hook.
Decay / dedupe / audit-edges / contradictions scans run on schedule; findings surface as [system] goals.
pnpm roushi <cmd> for local ops; pnpm mcp for stdio MCP server. Both share the same library code.
/api/mcp Route Handler exposes the same 35 tools to remote agents. Bearer-token auth via ROUSHI_MCP_HTTP_TOKEN.
Claude Code skills stay on disk for native /<skill> invocation; Roushi indexes them so think can recommend them.
Content in git is the source of truth. Neon Postgres is the index. Survives DB loss without losing knowledge.
Roushi is in private beta — currently single-tenant for KumoKodo's own portfolio of 20+ products. Multi-tenant SaaS rollout is on the roadmap once the brain primitives prove their value over a full quarter of dogfooding. Get in touch if you operate a 5+ product portfolio and want to be a design partner.