Y Combinator-backed Twill.ai moves AI coding assistants into isolated cloud sandboxes, letting developers delegate tasks through Slack or GitHub and wake up to finished pull requests.
Willy and Dan, the co-founders behind Twill.ai, ran into a frustration shared by growing numbers of developers working with AI coding tools. Claude Code and similar command-line assistants are powerful, but running them locally creates friction. Close your laptop and the agent stops. Try running two tasks that touch the same configuration files and you are stuck manually rebinding ports. Hand an autonomous agent full access to your local filesystem and you are making a significant bet on its reliability.
Their solution, launching now as part of Y Combinator's S25 batch, is straightforward: move the agents to the cloud. Twill spins up a dedicated sandbox for each task, clones your repository, installs dependencies, and invokes whichever coding CLI you select, whether that is Claude Code or another option. When the work is done, it comes back to you as a pull request, a code review, a diagnosis, or a clarifying question.
What makes Twill worth paying attention to is its architectural bet. Rather than building a proprietary AI coding harness, the platform reuses the native command-line tools that labs like Anthropic and OpenAI already ship. This matters because those labs are pouring enormous resources into reinforcement learning for their coding products. Every improvement Anthropic makes to Claude Code, for instance, automatically flows through to Twill users. The platform inherits the upside without maintaining its own model layer. Developers also avoid vendor lock-in: you can pick a different CLI for each task or combine them as needed.
Each task gets its own filesystem, its own ports, and full process isolation. Secrets are injected at runtime through environment variables and never persist in the sandbox itself. Once a task completes, Twill snapshots the filesystem so subsequent runs on the same repository start warm, with dependencies already installed. In practical terms, a three-person team can assign Twill a Linear backlog ticket about adding a CSV import feature to a Rails app. Twill clones the repo, implements the feature, runs the test suite, takes screenshots, and opens a pull request. If the PR needs revision, the team requests changes through GitHub, and Twill handles the next iteration.
For more complex tasks, the agent asks clarifying questions before writing code and records a browser session video as proof of work. Teams can also set standing instructions, such as always using an existing logger rather than console.log, that carry across every future task. There is also support for cron-based scheduling for recurring work and event triggers for situations like broken CI pipelines.
A competitive and crowded market
Twill enters a space packed with well-funded competitors. AI labs ship their own coding products directly. Integrated development environments wrap large language models inside editors. A growing wave of startups builds custom cloud agents on bespoke harnesses. The broader AI-assisted coding market is expanding rapidly, with tools like GitHub Copilot already generating significant revenue and adoption, so the appetite for solutions that reduce friction in developer workflows is clearly there.
The distinction Twill is drawing is that it does not compete with the labs on model quality or harness design. It competes on infrastructure and orchestration. By open-sourcing agentbox-sdk, an SDK for running and interacting with agent CLIs across sandbox providers, the founders are also signalling that they want to build a layer the ecosystem can standardise on, rather than locking users into a proprietary runtime.
Pricing starts with a free tier of 10 credits per month, where one credit equals roughly one dollar of AI compute at cost with no markup. Paid plans begin at $50 per month for 50 credits, with bring-your-own-key support on higher tiers. Open-source projects get free pro access. As the Financial Times recently noted, developer tools that reduce context-switching and automate repetitive coding tasks are becoming central to how engineering teams think about productivity. The question for Twill is whether orchestrating existing CLIs in sandboxes is enough to differentiate, or whether the labs will eventually bundle this kind of cloud infrastructure themselves. For now, the bet on lab-native tools and portable infrastructure looks pragmatic. If Anthropic, OpenAI, or others ship a better coding CLI tomorrow, Twill users get the benefit immediately without migrating platforms. That agility, combined with genuine isolation and persistence, addresses the three walls the founders hit themselves. Whether it scales into a lasting business depends on execution, but the problem it solves is real and growing.
