Hoplite launches cloud platform for coding agents
Bence and Ryan, the two founders of Hoplite, a Y Combinator Summer 2026 (YC S26) startup, introduced their product on Hacker News through a Launch HN post. Hoplite deploys coding agents inside isolated cloud sandboxes connected to a GitHub repository. Each agent can read and edit code, run the project's test suite, start a live preview of the app, use a persistent browser to check its own work, and open a pull request.
During onboarding, Hoplite copies over a developer's local setup, including sessions, memories and MCP servers, so cloud threads mirror a local workflow. The founders said they built a custom in-house agent harness rather than wrapping an existing solution like Codex or Claude Code, in order to add new features without waiting on Anthropic or OpenAI. Bence said the harness is internally called phalanx.
On infrastructure, Hoplite runs on AWS for most services, with three exceptions: Temporal handles durable workflows, Modal runs the sandboxes, and Planetscale hosts the database. In the comments, Bence explained that Modal's autoscaling has to stop a machine, resize it and start it again, a process that takes about 3 seconds and kills every running process in the sandbox. Daytona, by contrast, can scale up without a restart, though not scale down; Hoplite currently keeps Daytona only as a fallback for a Modal outage, because Daytona imposes tight limits on how many resources it can consume concurrently.
Pricing drew the most pushback in the thread. Hoplite offers $100 in free credits through the promo code HACKERNEWS at launch. Commenter fishtoaster described the plan as $99 per seat per month and asked for a cheaper, pay-as-you-go tier for personal projects; the founders acknowledged the feedback but did not restate the figure themselves. Another commenter, mellosouls, compared that to GitHub Copilot Cloud and Codex Cloud, which he said start at $20 a month, and asked what Hoplite offers to justify the gap. Bence replied that Hoplite's higher price reflects not adding a markup on top of token or sandbox costs, and that the company deliberately avoids usage-based upcharges because it wants users running as many agents as they can handle.
On models, Bence said Hoplite is not locked into one provider and can offer open-weight models such as Kimi K3 and GLM 5.2 alongside proprietary options. Hoplite's own documentation, however, describes the choice as being between "Anthropic and OpenAI frontier models," a discrepancy the thread never resolved. The founders also framed the bigger bet behind the product: they expect that within 6 to 12 months, mainstream development will shift from reviewing individual lines of agent-written code toward reviewing the resulting product behavior, such as new user flows, visual correctness and API behavior, while running many agents concurrently.
Commenters raised other points: a naming collision with an unrelated existing product called Hoplite.gg, criticism that the landing page read like AI-generated marketing copy, and questions about how Hoplite differs from Cursor's cloud agents, GitHub Copilot Cloud, Codex Cloud and exe.dev. On the recurring problem of agents over-iterating on review comments, Bence said Hoplite counters it by keeping tasks narrowly scoped, though that approach works less well when a review comment touches a complex architectural question.
Key facts
- Hoplite (YC S26), founded by Bence and Ryan, launched a cloud platform that runs coding agents in isolated sandboxes tied to a GitHub repository, able to edit code, run tests, boot a live preview and open pull requests.
- The company built a custom in-house agent harness instead of wrapping Codex or Claude Code, aiming for independence from Anthropic's and OpenAI's release cycles; Bence said the harness is internally called phalanx.
- Infrastructure runs on AWS except for Temporal (workflows), Modal (sandboxes) and Planetscale (database); Modal's autoscaling takes about 3 seconds to stop, resize and restart a machine, killing running processes, so Daytona serves as a fallback despite its own concurrency limits.
- The launch promo code HACKERNEWS grants $100 in free credits; a commenter cited $99-per-seat-per-month pricing against a roughly $20-a-month base for GitHub Copilot Cloud and Codex Cloud, figures the founders did not independently confirm in the thread.
- Hoplite says it is not locked to one model provider and offers open-weight models like Kimi K3 and GLM 5.2, though its own documentation describes model choice as limited to "Anthropic and OpenAI frontier models."
Why it matters
The launch adds another entrant to the fast-growing market of cloud coding-agent platforms, alongside Cursor's cloud agents, GitHub Copilot Cloud, Codex Cloud, Amp and exe.dev. Hoplite's bet is that developers will soon spend less time reviewing agent-written code line by line and more time reviewing what the resulting product actually does, and it is pitching a custom-built agent harness and full sandboxed dev environments rather than a thin wrapper around existing coding assistants.
Who it affects
Developers and teams evaluating cloud-hosted coding agents, particularly those currently running self-hosted setups such as Claude Code behind Docker Compose, or companies building similar cloud-agent infrastructure in-house, who now have a commercial product to compare pricing and features against.
How to use it
New users can sign up at hoplite.sh, connect a GitHub repository, and enter the promo code HACKERNEWS for $100 in free credits; Hoplite also lets users connect an existing Codex or ChatGPT subscription to run OpenAI models through it. The founders did not state pricing directly in the thread; a commenter cited a $99-per-seat-per-month tier, higher than the roughly $20-a-month base level another commenter cited for GitHub Copilot Cloud and Codex Cloud.
How solid is it
The product is live, documented at hoplite.sh/docs, and backed by Y Combinator's Summer 2026 batch; the founders answered dozens of detailed technical questions in the thread about infrastructure choices, pricing rationale and model support. Some claims are softer: the $99/seat/mo and $20/month figures come from commenters' own descriptions rather than a pricing page quoted by the founders, and Hoplite's documentation describing model choice as limited to "Anthropic and OpenAI frontier models" was never reconciled with the founder's claim that Kimi K3 and GLM 5.2 are also available.
Risks and caveats
The source gives no revenue, user count or headcount figures, and no named customers or case studies appear, only opinions from HN commenters. One commenter who is building similar cloud-agent infrastructure in-house argued that porting an established team's development environment into the cloud is a "tarpit problem" that teams with existing infrastructure are unlikely to buy into. The product also faces a naming collision with an unrelated existing app called Hoplite, and at least two commenters said the launch's marketing copy read like AI-generated text.
“Our pricing is higher than other providers because we do not upcharge on token or sandbox costs.”
— Bence (BenceRed), co-founder of Hoplite