Lumen — AI Marketing OS & Agent Suite

Lumen is an end-to-end AI Marketing Operating System made of two integrated Next.js applications: Lumen AI Agent, a conversational marketing copilot, and Marketing OS, a production SaaS workspace.

Marketing OS provides campaign management, an AI Content Studio (11+ content types), a Cloudinary asset manager, a live landing page builder (/p/[slug]), gated lead magnets (/m/[slug]) with verified email capture and CRM lead tracking, and an editor for the product playbooks that ground the agent.

AI chat tools are good at marketing strategy but disconnected from where campaigns actually run. Plans stay in chat, while assets, pages and leads live in other tools. Pasting full brand and product context into every prompt also bloats the context window and dilutes answers.

Instead of one giant system prompt, a heuristic detector reads the chat history backwards — newest message first — looking for product and skill keywords. It stops at the first match, then loads that product's playbook files and the matching SKILL.md (from 30 skills such as SEO, ads, competitor analysis and product research).

The system prompt is assembled from a base persona plus just those files. Playbooks and skills are stored as rows in a Postgres markdown_files table, with product keys validated and file counts capped, so each turn carries only the context it needs.

When the agent needs structured input, it wraps a JSON question set in <mcq> tags. The frontend hides the raw text and renders an interactive <InteractiveMcq> form (radio buttons, text inputs) inside the streamed answer. Submitting it writes the answers back as the user's next message, so the conversation continues naturally.

Users pick a model in the UI. The router maps it to DeepSeek, OpenAI, Anthropic Claude or Google Gemini. If the chosen provider has no API key configured, it falls back through the available providers in order rather than failing the request.

The only paid AI dependency is the LLM itself; search and reading use free services. Search runs SearXNG and DuckDuckGo Instant Answer in parallel, uses DuckDuckGo Lite only if fewer than five results come back, and Wikipedia only when everything else returns nothing. Results are normalized, de-duplicated and stripped of tracking parameters.

Page reading uses Jina Reader with a local Readability + Turndown fallback and a circuit breaker for outages. Every outbound request goes through one safe-fetch layer: SSRF checks, DNS validation and pinning, per-hop redirect validation, timeouts and response-size limits. Site crawls respect robots.txt and stay same-origin.

Marketing OS owns users, workspaces and the playbook editor. When a playbook is created or updated there — including from an uploaded document — it is written to the markdown_files table, and the agent picks it up on the next message.

Sign-in carries across with single sign-on: Marketing OS issues a signed JWT that expires in 60 seconds, and every token ID is recorded when used so the same link can't be replayed. The agent can also accept logins from accounts created in Marketing OS by reading its user table over a separate, optional connection.