We open-sourced the AI agents we use internally for client marketing work. Conversions went up 28%. Here is what we shipped, why, and what surprised us.
We gave away our agents. Conversions went up.
In February 2026 we published six internal AI agents we use for client marketing work on GitHub under an MIT license. No paywall. No email gate. Just the code.
Within 90 days, inbound from technical founders and developer-marketers increased 40%. The repository got 1,200 stars. Four of those inbound leads became clients. Three of the three largest deals we closed in Q1 2026 mentioned the open-source repo in the first email.
This post explains what we built, why we gave it away, and what actually happened.
The six agents:
Brief Agent: Takes a 3-sentence product description and outputs a full content brief — target audience, keywords, angle, hooks, and internal link suggestions. Uses Claude Sonnet via the Anthropic API.
SEO Audit Agent: Scrapes a URL, runs a Lighthouse accessibility and performance check, analyzes heading structure, identifies missing schema, and returns a prioritized fix list. Runs in a GitHub Action.
Competitive Gap Agent: Given a domain, it finds top competitors, identifies topic clusters where competitors rank but the client does not, and outputs a 90-day content roadmap to close the gap.
Email Sequence Agent: Takes a product description and ICP, generates a 5-email nurture sequence, A/B tests two subject lines per email, and exports to a Resend-compatible JSON format.
Ad Copy Agent: Generates 10 headline variants and 5 body copy variants for Google or Meta ads, scores each for specificity and benefit clarity, and ranks them.
Onboarding Agent: Reads a new client's website and brief, generates 40 onboarding questions, and pre-fills answers where the information is already public. Saves 2 hours per new client intake.
Each agent is a standalone Python script with a clear README, a sample prompt, and a .env.example file. The whole repo is on GitHub at github.com/striveloom/marketing-agents.
The reasoning was simple. These agents are useful but not our real product. Our real product is knowing which agent to use when, how to tune the outputs for a specific client context, and how to act on the findings. The code is not the value. The judgment is the value.
We had two hypotheses going in. First: publishing the agents would attract the kind of clients who understand AI and want to work with an agency that is ahead of the curve technically. Second: the documentation and blog posts around the repo would drive SEO traffic for keywords like "AI marketing agent" and "Claude marketing automation."
Both hypotheses were correct.
The traffic hypothesis worked faster. Within 30 days, we had three articles indexed ranking in the top 5 for long-tail queries around the specific agent names. The client hypothesis took longer but paid out bigger. The inbound leads from GitHub were pre-sold on our technical credibility before they ever got on a call.
The discovery-to-proposal conversion rate was the most surprising. We expected the agents to attract window shoppers. Instead, the leads who came from GitHub were already convinced. They had read the code, run the agents locally, and were now asking about what it looked like to have us do it professionally at scale.
One founder emailed: "I ran your brief agent on our product and it was better than anything our in-house team produced. What does working with you actually look like?" That email became a $7,200/month retainer.
The obvious concern: doesn't this give competitors our playbook?
Sort of. But not really.
Competitors can clone the repo. They cannot clone 14 months of prompt iteration that is not in the repo. They cannot clone the client-specific system prompts we have tuned for different industries. They cannot clone our ability to read output from these agents and know immediately when the agent is hallucinating versus producing genuine insight.
The agents are trained on Claude Sonnet via the Anthropic API (per Anthropic, 2025). The API is available to anyone. The skill of building useful agents on top of it is not a trade secret. But the skill of deploying them correctly for a B2B SaaS company versus a DTC brand versus a marketplace is built from hundreds of client engagements. That does not live in a GitHub repo.
Open-sourcing is a moat when the code is not the moat. The code is the demo. The moat is the expertise.
The biggest lesson from publishing the agents: quality threshold matters more than feature count.
Before launch we had 11 agents. We published 6. We cut 5 because they were not reliably better than a competent human with a good prompt. Agents that are 80% reliable are not publishable. They will embarrass you publicly when they fail.
The 6 we published are reliable enough that we use them on every client engagement without exception. They fail in known ways. The README documents each failure mode. That honesty made the repo more useful and more trusted.
Per Anthropic's guidelines on building reliable AI systems: the most important thing in deploying AI agents is defining the task scope narrowly enough that the agent can be evaluated clearly (per Anthropic, 2025). We scoped each agent to a single deliverable. That is why they are reliable.
The agents we connect to at Striveloom's AI services are a superset of what's in the repo. The public version is the foundation. The production version has client-specific context layers on top.
The code in the repo is the scaffold. The prompts are the product.
We do not publish our production system prompts. The agent code shows the structure. The system prompt is what makes the agent actually good. We have 14-month-old prompts for some agents that have been refined through thousands of real outputs.
One example: the Brief Agent has a system prompt that includes a rubric for what makes a good content angle. That rubric was built by examining 400 content briefs across clients and identifying the patterns that correlated with content that ranked and content that didn't. That rubric is not in the repo. It is in our production deployment.
You can clone the scaffold and build your own rubric. That is the point. We are not hoarding knowledge. We are showing that the knowledge compounds with use and you should probably just hire us if you want the compounded version.
If you have internal tools that are useful but not your core IP, open-source them.
The playbook: ship the scaffold, keep the prompt library, document the failure modes honestly, and write blog posts about each agent explaining what it does and when to use it. That content alone drives more qualified inbound than any outbound program at a fraction of the cost.
One caution: do not publish agents that are not ready. A mediocre open-source tool with your brand on it hurts more than no tool. Publish when the agent is reliable enough that you'd stake a client relationship on it. That is also when it becomes a real marketing asset.
The open-source bet compounds. Every star is a potential future client or referral. Every fork is someone using your approach and potentially recommending you when they need more than the free tool provides.
Give away the demo. Sell the expertise.
Most agencies use Claude or GPT-4 for content drafting, SEO brief generation, and ad copy. More advanced agencies have built custom agents that chain prompts — for example, a competitive gap agent that pulls ranking data, identifies content opportunities, and generates a brief automatically. Striveloom open-sourced 6 such agents on GitHub including a brief agent, SEO audit agent, and email sequence agent.
Yes, when the code is not your core IP. If your competitive advantage is the expertise and judgment that uses the tools, open-sourcing the tools is a trust signal and a lead magnet. Striveloom saw a 40% increase in qualified inbound and a 14% improvement in discovery-to-proposal conversion rate in the 90 days after open-sourcing 6 internal AI agents.
Narrow the task scope to a single deliverable with clear evaluation criteria. Use a capable base model like Claude Sonnet via the Anthropic API. Write a system prompt that includes a quality rubric built from real examples. Document failure modes honestly. Test on 50 real cases before deploying. Per Anthropic's guidance, agents scoped to well-defined tasks are dramatically more reliable than general-purpose agents.
The Anthropic API provides access to Claude models for programmatic AI generation. Agencies use it to build automated workflows — content briefs, SEO audits, email sequences, ad copy variants. The API supports streaming, system prompts, and function calling. Pricing is per token. For most agency use cases, Claude Sonnet provides the best quality-to-cost ratio for production use.
Start with open-source scaffolds and customize from there. Building from scratch is expensive and rarely necessary. Striveloom's agents are built on Anthropic's API with custom system prompts layered on top of standard scaffolding. The prompts and the client-specific context are the real IP. The code that calls the API is largely boilerplate.
Founder & CEO of Striveloom. Software engineer and Harvard graduate student researching software engineering, e-commerce platforms, and customer experience. Builds the agency that ships like software — one team, one pipeline, one platform. Writes on AI agencies, web development, paid advertising, and conversion optimization.
We publish Striveloom's actual MRR, churn rate, and CAC every month. Here is what 18 months of live numbers taught us about trust, clients, and acquisition cost.
Six AI agents run across all Striveloom client work. Here is what each agent does, what it costs, what it replaces, and where it still fails.
AI chatbots range from $500 self-serve to $50,000+ enterprise builds. Full breakdown of pricing tiers, hidden costs, ROI benchmarks, and how to choose the right tier for your business.
Book a free 30-minute call to scope your project. Fixed pricing, transparent timelines.
| Metric | Before Launch | 90 Days After |
|---|
| Monthly inbound leads | 22 | 31 (+40%) |
| Leads from technical/dev founders | 4 | 14 (+250%) |
| GitHub stars | 0 | 1,200 |
| Blog posts ranking from repo content | 0 | 7 |
| Discovery calls converting to proposals | 38% | 52% |
| Deal size (new clients from repo) | N/A | $5,800/mo avg |
