Marketing OS, SourceForge's multi-tenant SaaS platform that runs a digital marketing agency end-to-end, now ships with more than 60 production AI agents across eleven functional categories — sales, SEO, content, paid media, social, email, WhatsApp, analytics, intelligence, orchestration and onboarding.
Each agent is defined in code using a Zod-typed input and output schema, a versioned system prompt loaded from a markdown file, an explicit Claude model tier (Haiku, Sonnet or Opus), a whitelisted tool set and explicit budgets for tokens, latency and cost. Every run is logged to an agent_runs observability table with full tracking of cached versus uncached tokens, latency, status, tool calls and a trace ID.
The platform's cost discipline rests on four levers — prompt caching on system prompts and brand context (with cache hit rates consistently above 80% in production), Anthropic's Batch API for non-realtime work (50% off list pricing), an enforced model mix of approximately 15% Haiku, 75% Sonnet and 10% Opus, and per-agent and per-client cost ceilings that hard-stop runaway runs.
Marketing OS supports three operating models — single D2C brand running themselves, marketing agency running many client workspaces under one operator, and managed-by-SourceForge where we operate the platform on the customer's behalf. Multi-tenancy is enforced through Postgres Row-Level Security with separate "app" and "system" database roles for web and worker pools respectively.
The platform is currently in production with D2C consumer brands across beauty, health, wellness, fashion and food-and-beverage verticals, and with two pilot marketing agencies running multi-client deployments. Billing is in INR via Razorpay or USD via Stripe, with workspace currency locked once billing starts.