Daytona

Daytona provides programmatic, stateful sandboxes for AI agents to execute code, use computers, and run reinforcement-learning workloads. CEO and co-founder Ivan Burazin described the product as “composable computers for AI agents”: environments accessed through an API that can retain work across pauses and subsequent sessions. The company pivoted from human development environments to agent sandboxes in January 2025.[1][3]

The central product distinction is persistent, configurable compute. An agent can start a sandbox, fork parallel branches to explore different paths, and snapshot its environment to preserve state across failures. In the interview published on May 21, 2026, Burazin explained the design as combining rapid startup with a computer that remains available until its work is finished—much as a person expects to close and reopen a laptop without losing its state.[1][3]

Daytona's earlier development-environment work supplies context for that shift: Burazin and CTO and co-founder Vedran Jukic had previously worked on Codeanywhere, and Burazin said experience building its infrastructure later informed Daytona's scheduler.[2][1]

Before Daytona, Burazin and his co-founder worked on server infrastructure and virtualization, then Codeanywhere, a browser-based development environment. Their subsequent work included the Shift developer conference; Burazin also served as Chief Developer Experience Officer at Infobip. In his retrospective account, Codeanywhere required them to build much of the stack themselves, giving the team practical experience it brought into Daytona.[1][2]

Daytona's original product was a self-hosted Development Environment Manager. Its open-source project launched on March 6, 2024 under Apache 2.0. Developers could run daytona create to set up a full environment locally or on a remote server. The company reported more than 2,000 GitHub stars in the first 48 hours—a measure of repository interest—and said its ARR had nearly doubled in the six months following its November 2023 announcement, without giving an absolute revenue figure.[2]

On June 25, 2024, Daytona announced it had secured an additional $5 million in seed funding after a $2 million pre-seed round. Upfront Ventures led the seed round, existing investor 500 Global participated, and Upfront's Kevin Zhang was announced as joining the board. The announcement identified Burazin and Jukic as co-founders and Goran Draganic as Chief Architect.[2]

Burazin traced the pivot to a late-2024 experiment combining Daytona's development infrastructure with OpenDevin, later renamed OpenHands, as a SaaS product. He recalled limited use of that service but substantial interest from people building agents who needed a compute runtime. Roughly 20–30 prospective users rejected the existing infrastructure. Their response showed the team that human development environments did not meet those agents' requirements.[1]

On New Year's Eve at the transition from 2024 to 2025, Burazin built an initial MVP and sent it to his CTO and co-founder, Jukic. Burazin recalled that Jukic liked the idea but rebuilt the implementation over two weeks. When the team demonstrated the rebuilt version to earlier prospects, users followed up asking, “Where is my API key?” Burazin described that demand as the point at which Daytona committed to the new product. Its first version was a Linux environment with configurable CPU, RAM, and disk; the broader composable-computer vision developed afterward.[1][2]

The sandbox product ran on bare metal from the outset. Burazin attributed its startup speed to Daytona's own scheduler, local CPU, RAM, and disk, and templates or snapshots preloaded on NVMe drives. Requests are directed to a machine holding the required snapshot, reducing the need to fetch state over a network. Daytona used colocation providers; Burazin said operating its own data centers did not then justify the capital and risk for the expected margin improvement.[1]

In the interview published in May 2026, Burazin reported approximately 60 milliseconds for a single sandbox's request-to-reply startup, including network latency, and about 75 seconds to start 50,000 sandboxes concurrently. He said Daytona's largest customer ran about 850,000 sandboxes daily. These were his reported measurements; the interview supplied no reproducible test protocol. He separately mentioned a request for 500,000 concurrent CPUs, which represented requested capacity rather than a delivered deployment.[1]

Burazin distinguished background agents, whose demand tends to follow daily and weekly human activity, from RL and evaluation runs that abruptly consume large allocations and then stop. He reported average utilization of 15%, with peaks around 90%, and said Daytona used customer capacity commitments to provision for those peaks. Fast CPU-environment startup also matters to training economics: delays between runs can leave more expensive GPUs waiting.[1]

For research workloads, Daytona's declarative image builder assembles requested dependencies and environment variables, propagates a snapshot, and lets users start many sandboxes from it. Burazin also described dynamic resource resizing and default Docker containers hardened with Sysbox, with nested Docker and K3S workloads supported. Burazin credited Alex and his team with bringing Daytona into RL and evaluation workloads. In the interview, the interviewer described Harbor as having adopted Daytona, helping its expansion into research workloads.[1]

Computer use became another product investment. Burazin described Windows sandboxes with snapshots and forking, while saying the computer-use effort had not yet reached general availability. macOS and GPU sandboxes remained forthcoming in that account. His own example was preparing a board report: after API-based integrations returned incomplete data, he used an OpenClaw agent in a Mac Mini virtual sandbox with read-only company accounts to log into websites and export the missing information.[1]

On February 5, 2026, Daytona announced a $24 million Series A led by FirstMark Capital. Participants included Pace Capital, Upfront Ventures, E2VC, Darkmode, Datadog, and Figma Ventures; FirstMark's Matt Turck joined the board. Daytona said it was hardware-constrained and planned to use the funding to add capacity, expand into new regions, hire, and support sales and marketing.[3]

The February release reported reaching a $1 million forward revenue run rate in under three months and doubling it six weeks later; it did not specify calendar dates for those milestones. That measure is distinct from recognized annual revenue or contracted ARR. Daytona named LangChain, Turing, Writer, and SambaNova as customers and reported a 20-person team. In the subsequently published May interview, Burazin reported 25 people, estimating that about 13 had worked with the founders for seven years or more. He identified rapid team responsiveness, especially through Slack, as a reason customers stayed.[3][1]

Daytona's public-code strategy changed across product generations. The development-environment core used Apache 2.0; Burazin said the sandbox product adopted AGPL 3. A later notice in the public repository states that core development moved to a private codebase as of June 2026 and that the repository would receive no further updates, fixes, or releases. Existing public code remained available to use, fork, and build on under its license, without support or warranty.[2][1][4]