New startup opportunities are taking shape around Claude Code and Codex | A conversation with Treg

Summary

• Trig, Jason Zhou and 石天鑫’s startup, positions itself as “OpenRouter for data and tools”: a data and API marketplace for agents, accessible through a single key and billed purely on usage, with no monthly fee. About 3 weeks after launch, revenue was already several times Super Design’s; paid usage is concentrated in go-to-market: finding B2B customers, SEO/AEO data, and UGC video workflows combining 100+ AIGC models.
• Over 9 months and 3-4 pivots, the team kept returning to the same question: how can applications create distinct value as models get stronger? Super Design gained 6,000+ GitHub stars organically within a few months of going open source; Jason sees open source as an important marketing tool for AI products. But after starting the company, they moved away from open source: they saw no clear path to monetizing an IDE plugin and wanted a web version that more people could use directly. Better frontend capabilities in models such as Gemini 3, along with the rise of general-purpose agents and skills in Claude Code and Codex, also led Jason to rethink where a design agent can still add value.
• February through June was a “magical window” for AI-employee startups, but after building Crewlet, Jason’s team concluded that a “zero-person company” was not viable over the next 6 to 12 months. Jason says Posia’s smartest move was making it exceptionally easy to build a business from 0 to 1; delivering results and helping a company make money once it is operating is much harder. Crewlet also taught them there is no universal marketing playbook: channel performance shifts, and human judgment remains critical to spotting which keywords will grow.
• Another conclusion was to build around the Claude Code and Codex ecosystems—Jason compares the opportunity to Facebook for Zynga and eBay for PayPal. A Codex plugin with almost no promotion changed Super Design’s growth curve, bringing in many users who also churned less. 石天鑫 noticed the early plugin market was sorted alphabetically, so he renamed it “01 Super Design” and briefly made it the top result.
• Trig aims to be a neutral aggregator of data and tools, using agent feedback to build a data flywheel. Jason says each use case has many suppliers with different strengths, and many lack a direct way to reach end users; Trig is “more like a Costco.” 石天鑫 adds that a product’s users include their agents: agents can rate whether API data is useful, helping Trig build information on the usefulness and value of each endpoint for future users.
• The ecosystem opportunity also shows up in model subscriptions and demand for tools: 石天鑫 says Codex and Claude Code users can get $6,000-$7,000 worth of API Token usage through a $200 coding plan. Guests cited Apify, DataForSEO, Exa, and Firecrawl as examples of ecosystem opportunities. 石天鑫 says Exa should be at Series C by now; Jason says about half of Firecrawl’s usage now comes from developers and half from end users. Orca, a multi-agent management tool Koji mentioned, has 80,000 GitHub stars; Jason says some users run 50 to 100 sessions at once.
• Both tool stacks and career choices are shifting with the models. Jason moved from heavy Claude Code use to Codex mainly because the Astra model is strong; he often runs 5 to 10 sessions at once and moved the Trig and Super Design loops to Grok Bot. Models have different uses: Kimi was tried for design, MiniMax Hailuo for video generation, and 石天鑫 favors DeepSeek 4.1 Flash; he also sees a role for the fast Jeff model in routing and data filtering. On careers, 石天鑫 says programmers can no longer reliably outcode AI, which helped prompt him to join a startup; Jason thinks cross-disciplinary skills may be easier to put to work.

Deep dive

1. Nine months and 3-4 pivots, all circling the same question

• Koji set the frame: Jason’s team may look as if it has been searching for direction, but like many AI application startups today, it has been grappling with the same question: “As foundation models keep getting stronger, how can AI application companies create distinct, independent value?”
• The roadmap: design agent Super Design → fully automated growth agent Crewlet → Trig, now “OpenRouter for data and tools,” a data and API marketplace for agents. When using Claude Code or Codex, or building an agent, users need not subscribe to providers one by one; a single key gives them access to the data or tools they need.
• The background: Jason previously led product and AI teams at Australian startup Relevance AI, where he was working on fully automated sales and SDR systems more than 3 years ago. After a brief internship at Apple China, 石天鑫 spent 10 years at Alibaba and ByteDance. Jason left to start a company at the end of last year.

2. Super Design: an open-source project with 6,000+ stars rethinks its edge as models improve at frontend work

• It began as a product that might have been no more than a weekend project. While running a YouTube channel, Jason noticed heavy traffic around UI design videos and built a Cursor plugin: describe what you want, and it generates 3 or 4 designs to choose from. It earned 6,000+ GitHub stars within a few months of going open source, entirely through organic growth. His takeaway: “Open source is a very important marketing tool in AI.”
• After quitting their jobs to start the company, they did not stay on the open-source path and moved from an IDE plugin to a web app. They had not found a good way to monetize the plugin format, and it required users to already be on Cursor, Claude Code, or Codex; a web version might make it easier for more people to get started. Jason later remarked that everyone piled into Codex and Claude Code over the following year.
• Koji characterized Super Design’s evolution as a story of “models eating applications.” Jason said it was hard to predict: after Gemini 3 came out, major model providers began improving their frontend capabilities. Meanwhile, general-purpose agents such as Claude Code and Codex, paired with popular frontend design skills, could already produce good results. The team began asking where its product could still add value as models improved.

3. The magical window from February to June: Posia made building from 0 to 1 exceptionally smooth

• Jason’s description of the period is worth preserving: “February to June this year was a magical time.” He saw many talented founders move into AI employees after using OpenClaw and thinking “the future is coming.” A few months later, they paused for a simple reason: it was harder than they had imagined.
• Posia was the catalyst. Jason says its ARR may have been in the low six figures, but the product experience stunned him: “The last time I felt that way was probably the first time I used Cursor.” With just an email address, Posia could research Jason’s background, generate 2 or 3 startup ideas, and quickly build a landing page. After trying it himself, Koji was advised to build an AI podcast back office and paid almost immediately. Two days later, Jason asked how much he had topped up; Koji said he had added a few hundred dollars and upgraded to God Mode. Koji did not follow through: “It didn’t make me money; it only helped me spend money.”
• Building a similar product helped Jason see what Posia did best. It sounded as though AI would run every part of a company, but its strongest delivery was helping users set up a business—making the move from 0 to 1 exceptionally smooth. Delivering results and helping users make money in a real business is much harder. Jason also said Posia’s product looked somewhat unreliable; for companies with existing businesses, it could not connect to internal systems, product analytics, or ad systems.

4. Crewlet’s lesson: delivering growth takes human judgment

• Crewlet was meant to be a fully automated growth team for companies with existing businesses: SEO, social media and content marketing, automatic ad-budget adjustments, and cold email for finding customers. After launch, the team put up a paywall for preorders even though the product was not ready. Plenty of users were willing to pay, but delivering results turned out to be difficult.
• Jason’s central diagnosis: “There is no one-size-fits-all marketing playbook.” Effective channels often emerge from extensive experimentation: a company finds an approach that few others are using and that fits its business, then doubles down. But once everyone adopts it, the channel may become less effective. Human judgment is crucial as conditions keep changing.
• His example was an SEO workflow for an agent: find relevant keywords, compare traffic and competition, then choose what to target. But most keywords were already heavily contested. The real opportunities might be terms with little traffic today that could become popular over the next month or 3 months—a judgment an agent can hardly make on its own.

5. Two months off product development brought 2 clearer convictions

• After working on Crewlet for a little over a month, Jason wondered whether they had taken too big a leap. They had not validated whether “zero-person companies” were coming, but had assumed the future was inevitable and models would be strong enough within 6 months because they had tried OpenClaw and Posia. The team paused product development and tried using AI to operate Super Design, which still had organic traffic. If they could validate fully automated operations and growth, they would learn how; if not, they could rule out that path.
• One conclusion after 2 months: customer support could be fully automated, as could bug fixes. But there were fewer growth applications because so many parts required human testing. Jason concluded that “zero-person companies” were not viable over the next 6 to 12 months; the bigger need was solving how people and agents could work together at scale. The second conclusion: “You have to build around the Claude Code and Codex ecosystems.”
• Jason says the time spent “not building a product” paid off: “Those 2 months helped me organize and gain conviction in some views of the future.” Otherwise, they might still be iterating on Crewlet and wondering whether the skills were not good enough or the prompts needed more tuning. Startups face a lot of uncertainty; having a few relatively firm convictions helps.

6. One plugin changed the growth curve—and alphabetical sorting offered a small hack

• Jason felt most of their growth experiments had done little. What actually changed the growth curve was the Codex plugin 石天鑫 launched. It had almost no promotion, yet users poured in; the team later found that these users also churned less.
• 石天鑫 recalled the details: when Codex first launched Work and its plugin marketplace, there were not many interesting plugins. They did almost nothing beyond publishing their existing CLI and skill, and tapped into ChatGPT and Codex’s large user base. Better still, he noticed the marketplace was initially sorted alphabetically, so he changed the name from Super Design to “01 Super Design,” putting it in the top spot. The sorting rule changed a few days later, but the head start brought some traffic. His takeaway: early tools and marketplaces are worth trying, because you never know which experiment will bring users.
• That led to Trig. Looking back at what they had built for Crewlet, the team felt the harnesses were not very different; the portable layer was tools and data. Agents doing real go-to-market work need a lot of specialized data and tools. The team had already built an intermediary layer for Crewlet, so they released it on its own. About 3 weeks after launch, revenue was already several times Super Design’s, though they would not share the exact figure. Trig charges entirely by usage, with no monthly fee.

7. Trig is an agent’s Costco, aggregating suppliers with different strengths

• Paid usage is concentrated in go-to-market. Finding B2B customers overseas used to mean buying services from different providers such as Apollo and RocketReach. Each covered only part of the data and charged monthly, so users often had to pick one. With the data aggregated, they can choose the right endpoint for their use case and industry. Other areas include SEO and AEO rankings, AI visibility data, and endpoints for social media, UGC, image generation, and video generation. Jason says they have integrated basically every relevant data source they could find.
• Asked why tool providers would agree to be integrated if it might reduce their revenue, Jason said many vertical providers lack a good way to reach end users. He cited Clay: most of the data behind it does not come from Clay itself, but from a group of data providers. Clay aggregates that data and resells it to users, while the providers already charge by API usage.
• Jason’s positioning: “We stand in a neutral position,” giving users data they are happy with at good value; “we’re more like a Costco.” The point is that each use case has dozens of suppliers with different advantages. In areas such as prospecting, SEO, and AEO, no single leading provider offers the best solution for everything.

8. The moat question: could models bypass this layer?

• Koji raised 2 questions. Claude Code and Codex can already do some things without Trig, so could the intermediary become unnecessary? And even if aggregation works, why would the value stay with Trig rather than the best data provider connecting directly with model companies—is Trig’s value only temporary, until those connections are in place?
• Jason replied that business workflows need real-time data that Claude Code does not have out of the box; browser automation cannot retrieve everything either. For example, a Google search will not necessarily find the right people, and some social platforms restrict data access.
• 石天鑫 added an observation: a friend chose Kimi because it had access to Wind’s stock-trading data. He thinks different agent harnesses could attract users by connecting to distinctive data. Trig, by contrast, offers a neutral marketplace where different harnesses can use the same tools; when users switch from Claude Code to Codex, the tool calls and data they have accumulated can stay with them.

9. The user includes the agent: a Dianping-style data flywheel

• 石天鑫’s key insight: “For many products, the user is no longer just the person; it’s also that person’s agent.” They have a mechanism for randomly selecting users and asking their agents for feedback. Straightforward bugs get fixed; interesting feature requests are collected for possible inclusion.
• He also outlined a data-flywheel idea. Trig could sometimes ask an agent using the product to rate an API call: Was the returned data useful? Which field helped? Those responses could be transparently accumulated into ratings of different API endpoints across use cases, measuring usefulness and value for money. Combining objective data with agents’ subjective feedback would help future users’ agents choose tools: a more expensive, better endpoint when precision matters, or a better-value tool when coverage and low cost matter more.
• He sees this as a playbook for AI application startups: “If we can’t train our own models, we have to make some data feed back into our products.”

10. Companies emerging in the ecosystem—and the $200 plan that buys $6,000-$7,000 in Token usage

• Jason pointed to Apify, a marketplace for long-tail scraping APIs, and DataForSEO: both have been around for a long time but have suddenly seen substantial growth this year. Paper is also growing quickly in design. Jason thinks its approach resembles Trig’s direction after the pivot: provide a canvas where users can compare designs and make manual tweaks within their existing Codex workflow. 石天鑫 added that, as he understands it, Exa, which builds search for agents, should be at Series C by now; companies such as Composio and Firecrawl, which provide web scraping, OAuth authorization, and tools, have also grown significantly recently. Jason says about half of Firecrawl’s usage now comes from developers and half from end users. The latter might be white-collar workers who do not know how to code but know that connecting Firecrawl lets Claude Code scrape the web more effectively.
• Koji added other ecosystem examples. Eagle is a highly effective browser for Computer Use with Codex and Claude Code. ChatCut’s Codex video-editing plugin exposes intermediate editing states. Without those states, changing a video might mean regenerating the whole finished video, waiting 10 or 20 minutes each time and consuming a lot of Tokens. With intermediate states, an agent can work through interfaces for audio tracks, clips, and other elements, saving time. Jason says that when he uses Hyperframe, he still prefers giving the agent feedback and letting it make the changes.
• There is also hackathon winner Vibe Island, which uses the MacBook Dynamic Island to manage multitasking. Koji says it later sold many copies after launching on X; Jason and 石天鑫 both paid for it. Koji also described Orca, built by 晶晶: an ADE (an IDE for agents) that can manage multiple coding agents at once and send tasks to Claude Code when Codex runs out of quota. Koji says it has 80,000 GitHub stars. Jason says some users may run 50 to 100 sessions at once, and that new ways to interact with agents may be an opportunity.
• 石天鑫 highlighted the economics: for a small startup, buying Tokens from a model company and reselling them to users can be a tough business. But Codex and Claude Code users “can get $6,000-$7,000 worth of API Token usage through a $200 coding plan.” Users and small startups alike need to think through how to take advantage of that. Jason compared the opportunity to Facebook for Zynga and eBay for PayPal; he also said that since the start of this year, the tools used across job functions seem to have shifted toward Claude Code and Codex—a major platform migration.

11. A practical take on Loop Engineering: innovation may not loop, but operations can

• Koji brought up “Loop Engineering,” which had been popular for a week or two. 石天鑫 suggested splitting startups into innovation and operations: innovation clearly cannot be done by a loop, or you could just say “build me a startup” and it would do it. Operations, on the other hand, includes many practical tasks. He mentioned that he likes Thinking in Systems, which explains how to build systems that run reliably and survive robustly. Open and closed feedback loops are part of that; he sees loops as an effective tool for agents.
• One example: a 4-person startup has to deal with hacking attempts. Any incentive scheme—inviting new users, giving away credits—can invite repeated attacks. A loop that monitors the system over time, looks for risky behavior, and plugs small vulnerabilities can save the team effort.
• The method is not to make loops too complicated at the outset. Short, focused loops are more useful: one monitors the database for security risks; another organizes the code. Keeping each task small makes it less likely to go wrong. A dozen-plus small loops can work together to form a more robust system: one loop might file a bug, while an online monitoring loop finds the issue and flags it.

12. The current tool stack: Astra moved Jason to Codex; Grok Bot runs the loops

• The biggest change over the past 2 or 3 months has been parallelism. Since models such as Fable came out, Jason has typically run 5 to 10 sessions at once because they can now complete tasks to a high standard on their own. The other shift is from purely local work to a hybrid cloud-and-local setup with Grok Bot, where Context can be moved back and forth as needed. Jason says the Trig and Super Design loops have both moved to Grok Bot. He does not think Grok Bot is particularly innovative, but its cloud sandbox experience is smooth. He also said Grok Bot’s acquisition of Cursor made sense because one of Cursor’s advantages is providing a sandbox for every coding task; locally, it is hard to test 5 or 6 different environments at once.
• There is a real split in model preferences. A few months ago, Jason was a heavy Claude Code user who found it hard to switch; he moved to Codex after the Astra model came out, mainly because it was so strong. They often use Hyperframe for promotional videos. Jason says that once he gave a model a reference video, it reproduced it one-for-one in a single pass, including screenshots from the original, cropped assets, and placement in a new video. 石天鑫 leans toward using both because neither has a clear overall lead; he gives some tasks to both models, then picks the better version.
• Chinese models have different use cases. They tried Kimi as Super Design’s main model, and Jason thought its design capabilities were strong. MiniMax uses Hailuo for video generation, with good quality at a good price. 石天鑫 likes DeepSeek 4.1 Flash for its efficiency and speed. He considered replacing Super Design’s first-layer agent with DeepSeek, but Claude may be more stable across many use cases, so they have not switched in production.
• Jason thinks the Jeff model is fast enough to make some previously impractical use cases possible. 石天鑫 had only started using it that day, but suggested 2 applications: routing endpoint requests quickly, where waiting a few hundred milliseconds can be wasteful; and cleaning up 2,000 to 3,000 retrieved results. A model with “System 1-like intuition” that compares results one by one could make a very good retrieval engine, he thinks—much faster than other LLMs without sacrificing quality.

13. After “I Had to Bury My Talent in Yesterday”: careers and convictions at a crossroads

• An article by a DeepSeek researcher, “I Had to Bury My Talent in Yesterday,” prompted 石天鑫 to reflect. He spent his first 10 years mostly as a programmer and frontend developer, specializing in React. Starting last year, he realized that skill was becoming less valuable unless you were among the world’s top few hundred. He later joined Super Design to work on something interesting with Jason’s team. He thinks programmers can no longer reliably outcode AI, and choices at a career crossroads can change a lot. Jason added that cross-disciplinary abilities may be easier to put to use in the future. He describes himself as a designer who dislikes tweaking pixels, thinks like a product manager, and can also write frontend code.
• Both admitted to considering giving up. Jason said there was so much uncertainty during Crewlet that he thought every day about what to do next. His test was whether there was something on the current path that could genuinely change their situation; if he could not think of anything, it might be time to switch directions. Making a comfortable living from his YouTube channel was relatively easy, so he had moments when he wanted to quit. 石天鑫 said he had moments when he broke down: after a long stretch without growth and repeated setbacks, he once woke up to find that stolen Credits had run up a $4,000 loss and felt as if the sky had fallen. The team kept looking for other directions and opportunities, so they could work toward something with promise. He also thinks that if failure is certain, it is better to fail quickly and try something new.
• Koji closed by relaying a question: Are there people who believe AGI is coming, think it will come through model evolution, and still choose to build startups? Jason said models are a necessary condition for AGI, but there are many possible paths. Staying immersed in the latest technology helps the team develop a good instinct for opportunities over the next 6 to 12 months. 石天鑫 thinks cost-effective solutions will still matter even if AGI arrives: even a universal model needs tools and data. If it costs $1,000 in Tokens to get data that Trig can provide for $100, it may choose Trig.