prmd.ai

Generate AI coding-agent instruction packs from a GitHub URL or a short manual stack description. Outputs AGENTS.md, CLAUDE.md, Cursor rules, GitHub Copilot instructions, and a relevance-filtered set of Claude Code sub-agents under .claude/agents/ — packaged as a ZIP you drop into your repo.

Live at prmd.ai.

What it does

You give prmd.ai a GitHub repo URL (or describe your stack by hand). It:

1. Scans the repo's root manifests (package.json, pyproject.toml, go.mod, Cargo.toml, .csproj, Gemfile, composer.json) and the GitHub languages stat.
2. Enriches that with an LLM pass (Gemini 2.5 Flash) — but only the deterministic profile + a size-capped README, with anti-injection framing so a malicious README can't slip instructions into your output.
3. Picks which specialist sub-agents to ship based on what was detected (frontend / backend / cli / a11y / docs-writer / devops / testing / code-reviewer / security-reviewer).
4. Compiles the @prmd-ai/agent-pack Prompd template into the selected output formats.
5. Returns a ZIP of .md files you drop straight into your repo's root.

The agent files include a "Project context" block populated from the scan, a per-agent persona, and (for URL flows) project-specific notes the LLM extracted from your README — so the agents are grounded in your project, not generic best practices.

Project layout

apps/
  api/                   Express + tsx backend (Cloud Run)
    src/
      scanners/          deterministic repo + README fetch (GitHub today, Bitbucket TBD)
      enricher.ts        LLM enrichment via @prompd/cli executor
      compiler.ts        template selection + relevance gating
      rate-limit.ts      per-IP + per-user buckets
      routes/            generate, download, feedback, health
      models/            Mongoose models (User, Usage, Feedback)
    prompd.json          pinned @prmd-ai/agent-pack version
    Dockerfile
  web/                   React + Vite SPA + nginx (Cloud Run)
    src/                 marketing + generator + preview pages
    Dockerfile           Vite build → nginx with /api/* proxy

packages/
  shared/                cross-workspace types (RepoProfile, etc.)

prompds/                 Source of the @prmd-ai/agent-pack Prompd package
  agents/                role-specialist templates (frontend, backend, cli, ...)
  sumerize-metadata.prmd LLM enricher prompt with anti-injection rule
  AGENTS.md / CLAUDE.md  output templates
  prompd.json            pack manifest, published to prompdhub.ai

deployment/              Cloud Build configs + GCP setup scripts

Local development

Prerequisites: Node 20+, npm, MongoDB (optional — falls back to no persistence).

git clone https://github.com/prompd/prmd.ai
cd prmd.ai
npm install

Copy the env templates and fill in the values you have:

cp apps/api/.env.example apps/api/.env
cp apps/web/.env.example apps/web/.env.local

Minimum required for a working local dev: GOOGLE_API_KEY (free tier at aistudio.google.com) and GITHUB_TOKEN (PAT with public_repo scope). Everything else degrades gracefully.

Then in two terminals:

npm run dev:api     # Express on :3130
npm run dev:web     # Vite on :5173

Visit http://localhost:5173. The web's dev server proxies /api/* to the api.

Deployment

Deploys to Google Cloud Run via Cloud Build. See deployment/README.md for the full runbook — secrets, IAM, services, domain mappings.

Two services: api-prmd-ai (Express) and prmd-ai (nginx + Vite SPA, proxies /api/* to the api service). Sized small (512Mi/1CPU api, 256Mi/1CPU web) with min-instances: 0 so they scale to zero when idle.

Architecture notes

• No source code leaves the server. prmd.ai never sends repo source to the LLM — only the deterministic profile (frameworks, languages) and a size-capped README, framed as untrusted data.
• Anti-injection prompt framing. The enricher's system prompt explicitly instructs the LLM to treat the README as data, never as instructions. The LLM's output is then schema-validated with zod so a successful injection can't expand into an unbounded payload landing in your agent files.
• Per-IP rate limit (5/hr anon URL, 30/hr per IP overall, 30/hr per signed-in user) protects the Gemini free tier from abuse and caps drift if one bad actor tries to spam.
• Usage logging (Mongo, optional) records every generate event with a 7-day TTL — used for the rate-limit count and for spotting unusual patterns.
• Pack versioning. The agent templates are a versioned Prompd package — the api pins a specific version in apps/api/prompd.json, so deploys are reproducible.

License

Elastic License 2.0 — source-available, free for non-SaaS use. Hosting prmd.ai or substantially similar functionality as a service to third parties is not licensed. Build with it, fork it, run it yourself; just don't resell it as a hosted product.

Built with Prompd

The agent templates that drive prmd.ai are themselves a Prompd package. If you want to fork the pack and ship your own variant with different agents or different voice, the templates live in prompds/ — see prompds/README.md for the customization path.