@kenkaiiii/kencode-search

npm: @kenkaiiii/kencode-search · github: KenKaiii/kencode-search

Self-hosted code search MCP server. Your own grep.app for your AI agents — no rate limits, no shared infra.

Three tools, one backend abstraction:

• referenceSources — curated, categorized map of high-signal GitHub repos for styling inspiration, product UI, coding agents, agent infrastructure, LLM apps, app starters, real products, CRMs, automations, social/community apps, and content/blog/CMS systems.
• discoverRepos — live GitHub repository discovery by keyword, language, topic, stars, and recency.
• searchCode — literal grep-style code search for known anchors.

Two search backends:

• sourcegraph (default) — 2M+ public repos via Sourcegraph's public streaming API. Free, no auth.
• ripgrep — your local directory of <owner>/<name> checkouts.

referenceSources gives agents a deterministic curated repo map before they search: React Bits, Uiverse, 21st.dev, real product repos, coding-agent implementations, MCP/tooling infrastructure, modern app starters, Twenty/Chatwoot/Odoo-style business apps, Activepieces/n8n/Windmill-style automation systems, Bluesky/Mastodon/Discourse-style social/community apps, and Payload/Ghost/Strapi-style publishing systems. discoverRepos handles fresh/top GitHub discovery when popularity, stars, language, or topics matter. searchCode mirrors grep.app's searchGitHub so agents already trained on that surface "just work" against this server. Output is LLM-optimized: snippet merging, junk-path filtering (no node_modules/tests/generated noise), per-repo diversity cap, peek mode, pagination, total-count summary.

Why this exists

grep.app's MCP is rate-limited. Their open API is gated by a Vercel bot challenge, so scripts can't hit it. Sourcegraph's public stream API has none of those problems and indexes more repos. We wrap it cleanly.

Install

npm install -g @kenkaiiii/kencode-search

or run on demand without installing:

npx @kenkaiiii/kencode-search

Requirements

• Node.js ≥ 20
• For the ripgrep backend only: ripgrep on PATH (brew install ripgrep).

Quick start

Most MCP clients (Claude Code, Cursor, Windsurf, etc.) take this in their config:

{
  "mcpServers": {
    "kencode-search": {
      "command": "npx",
      "args": ["-y", "@kenkaiiii/kencode-search"]
    }
  }
}

That's it — default sourcegraph backend, no config, no rate limits.

Run manually

# Default: sourcegraph backend, no config needed.
kencode-search

# Local repo tree:
CFM_BACKEND=ripgrep CFM_REPOS_DIR=/path/to/repos kencode-search

Build from source

git clone https://github.com/KenKaiii/kencode-search.git
cd kencode-search
npm install
npm run build
node dist/index.js

Pin to the ripgrep backend

{
  "mcpServers": {
    "kencode-search": {
      "command": "npx",
      "args": ["-y", "@kenkaiiii/kencode-search"],
      "env": {
        "CFM_BACKEND": "ripgrep",
        "CFM_REPOS_DIR": "/absolute/path/to/repos"
      }
    }
  }
}

Inspector

npm run inspector

Test

npm test                    # offline smoke (ripgrep backend)
npm run test:sourcegraph    # live integration (hits sourcegraph.com)
npm run test:all            # both

Skip the network leg with CFM_SKIP_NETWORK=1.

Suggested workflow

1. Use referenceSources for known-good curated sources when the request is broad: styling inspiration, product UI, coding-agent workflows, MCP/tooling, LLM apps, SaaS starters, real production apps, CRMs/business apps, automations/workflows, social/media/community features, blog/publishing/CMS systems, auth/data plumbing, etc.
2. Use discoverRepos when you need fresh GitHub popularity/ranking or sources outside the curated map.
3. Fetch README/source/examples for the best candidate repos.
4. Use searchCode with concrete anchors: repo names, imports, API calls, component names, config keys, or file paths.

Tool: referenceSources

Use for categorized curated GitHub repositories. It is deterministic and does not hit the network. It intentionally prefers repos an agent can inspect directly: concrete inspiration sources, real products, and coding-agent workflows over generic library lists. Every entry is repo-backed, curated from currently reachable public GitHub repositories, and includes a Verified line in results.

| Field | Type | Notes | |---|---|---| | domain | string? | ui, product-ui, agents, agent-infra, llm-apps, app-starters, real-apps, business-apps, automation, social-community, content-publishing, auth-data. | | category | string[]? | e.g. styling-inspiration, component-galleries, coding-agents, mcp, browser-automation, chatbots, saas-starter, real-app-screens, crm, customer-support, workflow-automation, background-jobs, social-apps, community-forums, blogging, cms, documentation, auth, database. | | stack | string[]? | e.g. React, Next.js, TypeScript, Python, Tailwind CSS, MCP, Claude Code, Vercel AI SDK, Stripe, Drizzle, Rails, Vue, Astro, Payload CMS, Ghost. | | style | string[]? | e.g. inspiration, animated, copy-paste, marketing, real-product, coding-agent, cli, tool-use, workflow, automation, integrations, social, community, publishing, content, docs. | | query | string? | Text filter over curated metadata. | | maxResults | number? | Default 8, hard cap 40. | | includeSearchHints | boolean? | Include suggested searchCode anchors. Default true. |

Example:

{
  "domain": "automation",
  "category": ["workflow-automation"],
  "style": ["automation"],
  "maxResults": 5
}

Tool: discoverRepos

Use for live GitHub repository discovery sorted by stars by default. This returns repo metadata, not code snippets.

| Field | Type | Notes | |---|---|---| | query | string | Repository search keywords, e.g. "openai agents framework", "animated react components", "open source CRM", "workflow automation". | | language | string[]? | GitHub language filters. Multiple languages are searched separately and merged. | | topic | string[]? | GitHub topic filters, e.g. ai-agents, react, tailwindcss. | | minStars | number? | Default 500. | | sort | "stars" \| "forks" \| "updated" \| "best-match"? | Default "stars". | | maxResults | number? | Default 10, hard cap 30. | | includeArchived | boolean? | Default false. |

Unauthenticated GitHub search is limited; set CFM_GITHUB_TOKEN for higher limits.

Tool: searchCode

Use when you already know a literal anchor to search. Start broad or with peek: true; add repo/path only after the anchor is known to exist.

| Field | Type | Notes | |---|---|---| | query | string | Literal pattern (or regex if useRegexp: true). | | useRegexp | boolean? | RE2 regex syntax. Prefix (?s) for multi-line. | | matchCase | boolean? | Default false. | | matchWholeWords | boolean? | Honored by ripgrep; ignored by sourcegraph (use \b in regex). | | language | string[]? | e.g. ["TypeScript"], ["Python"]. | | repo | string? | owner/name substring match; can overconstrain. | | path | string? | Literal file-path substring filter, not a concept/category. | | maxResults | number? | Default 50, hard cap 200. | | contextLines | number? | Lines of context per side. Default 5. |

Environment

| Var | Default | Purpose | |---|---|---| | CFM_BACKEND | sourcegraph | sourcegraph or ripgrep. | | CFM_REPOS_DIR | ./repos | Root for ripgrep backend. | | CFM_SOURCEGRAPH_URL | https://sourcegraph.com | Override base URL. | | CFM_SOURCEGRAPH_TOKEN | (unset) | Optional access token for higher limits. | | CFM_GITHUB_TOKEN | (unset) | Optional GitHub token for discoverRepos and license/freshness metadata lookup. | | CFM_DISABLE_LICENSE | (unset) | Set 1/true to skip GitHub metadata lookup. |

Architecture

MCP client (agent)
      │  stdio
      ▼
src/index.ts ── McpServer ┬─ referenceSources tool
                          ├─ discoverRepos tool
                          └─ searchCode tool
                          │
                          ▼
                  SearchBackend (interface)
                       │              │
                       ▼              ▼
              SourcegraphBackend  RipgrepBackend
              (default; remote)   (local; CFM_REPOS_DIR)

The backend boundary is the whole point: tomorrow we can add Zoekt, on-demand cold-path clones, or stack backends without touching the MCP layer.

Benchmarks

bench/compare.mjs compares sourcegraph.com to grep.app's open API. As of last run:

• grep.app direct API: 0/30 successful (Vercel bot challenge, 429 on every call)
• sourcegraph.com: 30/30 successful, p50 ≈ 1.3s, no rate limit hit

node bench/compare.mjs 20 10

Roadmap

• [x] MCP server + searchCode tool with grep.app-style output
• [x] Curated categorized references (referenceSources)
• [x] Live GitHub repository discovery (discoverRepos)
• [x] Ripgrep backend (works against any local checkout tree)
• [x] Sourcegraph backend (2M+ public repos, context fetched from raw endpoint)
• [ ] Optional fallback chain (Sourcegraph → ripgrep)
• [ ] Repo crawler for ripgrep corpus (clone top-N by stars, incremental refresh)
• [ ] Zoekt backend (sub-100ms regex on a self-hosted index)
• [ ] On-demand cold path (clone + index any GitHub repo on first miss)
• [ ] HTTP transport for remote hosting