Hoplite, a Y Combinator S26 company, has launched a platform for deploying cloud-based coding agents, giving each agent thread its own isolated real machine cloned from the user's repository. The product, announced via Y Combinator's launch page and a Hacker News post by founders Bence and Ryan, targets developers who want autonomous coding workflows running in the cloud rather than locally. (Hoplite)

Each sandbox can be configured with up to 8 vCPU, 32 GiB RAM, and 10 TB of disk. The platform supports models from Anthropic, OpenAI, and open-weights providers, selectable per thread. Users who already have a Codex subscription can connect it and route OpenAI models through Hoplite. (Hoplite)

The agent workflow follows a deliberate loop: it reads code, edits files, runs tests, inspects failures, retries, and pauses for human sign-off before opening a pull request. Every sandbox also boots the application on a live URL preview that both the user and the agent can interact with through a browser, giving the agent a feedback channel beyond terminal output. (Hoplite)

On the integration side, Hoplite connects to GitHub, Linear, Slack, and MCP. Its CLI can import local Claude Code or Codex history, including sessions, memories, and MCP server configurations, porting a developer's existing setup into the cloud environment. Users can spin up autonomous coding tasks manually from Slack, Linear, or iMessage, and the Automations feature supports triggering agents from webhooks and schedules. Sentry integration enables automatic task creation from error events. (Hoplite, Y Combinator Launches)

Pricing follows a prepaid credit model. Runs are metered individually, with each run displaying its cost. The Scale plan permits 150 concurrent sandboxes per seat. At launch, Hoplite offered a free trial with $100 in credits via the code "HACKERNEWS." (Hoplite, Hacker News)

Under the hood, Hoplite is hosted on AWS and uses Temporal for durable workflows, Modal for sandbox provisioning, and PlanetScale as its database. The founders built a custom in-house agent harness rather than relying on off-the-shelf solutions like Codex or Claude Code. (Hacker News)

The company did not start here. Bence and Ryan originally applied to YC with an AI product for retail investing but abandoned it because, as they wrote, it was not a product they themselves would use. The pivot to cloud coding agents produced Hoplite's current offering. (Hacker News)

The architecture decisions here are worth unpacking. Giving each agent thread a dedicated machine rather than a container or shared VM is a meaningful resource commitment, but it sidesteps the noisy-neighbor problems and security boundary concerns that come with shared environments. The 10 TB disk allocation per sandbox suggests the platform expects agents to work with large monorepos or data-heavy projects where container image pulls would introduce unacceptable latency. And the live URL preview, where the agent can interact with its own running application via browser, closes a gap that has limited local coding agents: without visual or runtime feedback, agents working on frontend or full-stack code are operating blind to the consequences of their changes.

The choice to build a custom harness rather than wrap Claude Code or Codex carries tradeoffs. Off-the-shelf harnesses provide faster time-to-market and benefit from upstream maintenance. A custom harness gives Hoplite control over the agent loop, retry logic, and integration points, but means the team owns the full burden of keeping pace with rapidly evolving model capabilities and API changes from multiple providers. The per-thread model selectability, spanning Anthropic, OpenAI, and open weights, makes that burden concrete: the harness must normalize across providers whose APIs, context windows, and tool-use conventions differ.

The concurrency story is aggressive for a new entrant. 150 concurrent sandboxes per seat on the Scale plan implies that a small team could run hundreds of parallel agent threads, each in its own isolated environment. Whether that capacity translates to productive work depends heavily on how well the agent harness manages coordination, deduplication of effort, and merge conflicts across threads operating on the same repository.

The pricing model, metered per run with visible costs, addresses a real pain point. Developers evaluating autonomous coding agents have had difficulty predicting spend, since agent loops are inherently variable in length. Showing cost per run, rather than billing on a flat subscription, lets teams build intuition for what specific types of tasks cost to complete. The prepaid credit structure, however, means teams must estimate their aggregate usage ahead of time.

The broader question Hoplite sits inside is whether the dominant locus of AI-assisted coding shifts from local IDEs to cloud-hosted agent environments. Local tools like Cursor and GitHub Copilot operate inside the developer's editor. Cloud-hosted agents can leverage compute resources that exceed what most developer laptops offer, run for longer durations without tying up local machines, and integrate more naturally with CI/CD-adjacent workflows. Hoplite's integration surface, spanning issue trackers, chat platforms, error monitoring, and source control, positions it as infrastructure for continuous, event-driven autonomous development rather than interactive pair programming. Whether that model wins depends on whether agent output quality has crossed the threshold where human review of pull requests is cheaper than human authorship, a bar that varies widely by task complexity and codebase maturity.