Rewrites coding-assistant prompts using strategies from awesome-agentic-patterns. Paste a prompt, PromptForge classifies the task, selects the best agentic patterns from a catalog of 22, and asks your configured LLM to rewrite it.

Also generates complete project repositories from a plain-language objective — paste a goal, get a downloadable .zip of a runnable project.

Available as a web app and an MCP server (stdio + HTTP). Same enhancer pipeline behind both surfaces.

• Vitest + Playwright suites green across 7 workspace packages — run make test for the current count.
• Typecheck + build clean.
• biome lint configured.
• CI workflow at .github/workflows/ci.yml.
• Architecture: see ARCHITECTURE.md. Doc-vs-code drift: see DRIFT.md (regenerated by /audit).

pnpm install                # bootstraps the workspace

Optional: install one of Ollama, Lemonade, or llama.cpp's llama-server locally if you want to drive the rewriter with a local model.

make dev
# or: pnpm --filter @prompt-forge/web dev

Open http://localhost:3000.

1. Click Settings in the header.
2. Pick a provider (ollama / lemonade / llamacpp / openai / anthropic), fill in baseUrl + model + (for openai/anthropic) apiKey, click Save.
3. Click Test connection under the saved provider — confirms the endpoint accepts a tiny "ping" chat.
4. Pick the active provider with the radio.
5. Back at /, paste your raw prompt and click Rewrite prompt.
6. Navigate to /repo to generate a full project from an objective.

Config is stored in localStorage under the key promptforge:config — nothing leaves your browser except the chat call to your configured provider.

Navigate to http://localhost:3000/repo. Enter a project objective (e.g., "simple Express.js REST API for todos with TypeScript"), click Generate. PromptForge sends a single LLM call and parses the returned file blocks into a downloadable .zip. No server-side storage — generation and zip assembly happen in the browser.

make dev-mcp-http
# or: pnpm --filter @prompt-forge/mcp start:http
# listens on http://127.0.0.1:8787/mcp (set PORT to override)

Send any JSON-RPC 2.0 request with content-type: application/json. Try:

curl -s -X POST http://127.0.0.1:8787/mcp \
  -H 'content-type: application/json' -d '{
    "jsonrpc":"2.0","id":1,"method":"tools/call",
    "params":{
      "name":"enhance_prompt",
      "arguments":{
        "rawPrompt":"refactor the auth module to use react hooks and add tests",
        "reflect":true,
        "provider":{"kind":"ollama","baseUrl":"http://localhost:11434","model":"llama3.1:8b"}
      }
    }
  }'

HTTP behavior:

• Methods: tools/list, tools/call, resources/list, resources/read, initialize
• OPTIONS /mcp returns CORS preflight (origin: * by default)
• POST body limit: 256 KiB (returns 413 on overflow)
• Notifications (no id) return 204 with empty body
• Streamable HTTP (MCP 2025-03-26 spec): POST with Accept: text/event-stream returns the response as a single SSE event. GET /mcp with the same Accept header opens an SSE channel with a Mcp-Session-Id header and keep-alives (default 15 s).
• Bearer auth (optional): set PROMPTFORGE_MCP_TOKEN=<secret> to require Authorization: Bearer <secret> on all POST/GET requests. Without the env var, no auth is required (back-compat). OPTIONS preflight is never gated.
• Bind address: set HOST=0.0.0.0 to expose on the LAN. Never do that without PROMPTFORGE_MCP_TOKEN set.

make dev-mcp-stdio
# or: pnpm --filter @prompt-forge/mcp start:stdio

Claude Desktop config:

{
  "mcpServers": {
    "promptforge": {
      "command": "pnpm",
      "args": ["--filter", "@prompt-forge/mcp", "exec", "tsx", "src/stdio.ts"],
      "cwd": "/absolute/path/to/this/repo"
    }
  }
}

Or, if you've installed the package globally (or via a bin path):

{
  "mcpServers": {
    "promptforge": {
      "command": "node",
      "args": ["/absolute/path/to/apps/mcp/bin/stdio.mjs"]
    }
  }
}

Server advertises { tools: {}, resources: {} }.

Input:

Returns text content with the rewritten prompt plus a header line listing the inferred taskKind and selected patterns.

Each of the 22 catalog patterns is exposed at promptforge://patterns/<slug> with mime type text/markdown. Use resources/list to enumerate, resources/read to fetch the directive.

LIVE_OLLAMA=1 make test-live
LIVE_OPENAI=1 OPENAI_API_KEY=sk-... make test-live
LIVE_ANTHROPIC=1 ANTHROPIC_API_KEY=sk-ant-... make test-live

Skipped silently when the matching LIVE_* env var is unset, so make test stays hermetic.

make test-e2e-install    # one time: pulls Playwright's chromium binary
make build               # e2e runs against the production build
make test-e2e

4 happy-path tests covering home, form gating, end-to-end rewrite (mocked /api/enhance), and the settings page.

prompt-forge/
├── apps/
│   ├── web/        Next.js 15 app (/ + /repo + /settings + /api/enhance + /api/repo + /api/providers/test)
│   └── mcp/        MCP server — JSON-RPC handler, stdio + HTTP transports, body limit, CORS
├── packages/
│   ├── core/       Result, AppError, Zod schemas, structured logger
│   ├── config/     AppConfig + memory/localStorage stores + pure helpers
│   ├── patterns/   22-entry catalog + classifier + selector
│   ├── providers/  ProviderClient: Ollama + OpenAI-compat + native Anthropic
│   ├── enhancer/   Pipeline: classify -> select -> chat -> (optional reflect) -> extract
│   └── repo-gen/   Single-pass repo generator: prompt builder + file-block parser
├── docs/           Architecture, packages, providers, patterns, TDD strategy
├── start.sh        One-shot launcher: web app + MCP HTTP server
└── setup-mcp.sh    Auto-detect AI tools and write MCP server config

1. Add the kind literal to ProviderKindSchema in packages/core/src/types.ts.
2. Create packages/providers/src/<kind>.ts exporting create<Kind>Client(config, fetchImpl?).
3. Route the kind in packages/providers/src/factory.ts.
4. Add a default base URL to DEFAULT_BASE_URLS.
5. Write the test with scriptedFetch.
6. Add the kind to PROVIDER_KINDS in apps/web/src/components/provider-settings.tsx.

Apache-2.0.