Railway

Railway is a cloud platform for deploying applications, servers and databases. Its AI-oriented products let coding agents deploy services, inspect logs and metrics, manage environments and control feature rollouts. In August 2026, it introduced dev.new, bringing app generation, previews, testing and deployment into one Railway project.[4][8][7]

In the interview published May 20, 2026, founder and CEO Jake Cooper said Railway had spent the preceding six months deeply prioritizing agents as a way to build and deploy software. His direction extended beyond accepting agent-generated code: he wanted agents to copy services and production data, test changes close to production, and submit a new infrastructure state for approval. These were ambitions for safer iteration, rather than a claim that autonomous production repair was already solved.[1][4]

Railway’s agent-facing tools build on an earlier deployment platform. Its history begins in 2020, before the agent strategy, with a product designed to deploy code and databases without requiring engineers to manage servers.[2][1]

Cooper’s earlier work included front-end development on Wolfram Mathematica, a brief period at Bloomberg and distributed systems at Uber. He recalled moving Jump Bikes systems onto Cadence, the workflow engine that preceded Temporal. Railway started in 2020 and was released in 2021, offering operations such as creating a Postgres instance or deploying code from GitHub, with defaults that users could adjust later.[1][2]

The canvas made infrastructure visual: users added services, and Railway reconciled their actions by deploying and connecting those services underneath. Early customer acquisition was personal. Cooper recalled keeping Discord open on a second monitor and immediately greeting newcomers while trying to get the first 100 users to return.[2][1]

On May 31, 2022, Railway announced a $20 million Series A, led by Erica Brescia and Jordan Segall at Redpoint Ventures. Seed lead Unusual Ventures and pre-seed lead Lachy Groom (LGF) also participated, bringing reported cumulative funding to just under $25 million. Its plans included expanding the canvas and using Nix for reproducible builds without user-written Dockerfiles.[2]

Free compute drove a surge in use during 2022–2023, but the economics forced a retreat. Cooper recalled losses of about $500,000 a month during the free-tier era and a subsequent two-year effort to make the business work. In its later retrospective, Railway reported that, at its worst point, it lost $16 for each $1 of topline revenue. It deleted the free plan and accepted a sharp drop in growth while rebuilding.[1][6]

Railway initially ran on Google Cloud Platform. The rebuild moved workloads to Railway Metal, with Railway building its own orchestrator, compute stack and edge network. Its August 27, 2025 retrospective credited the move to owned servers with changing its margins, while noting that builds still ran in the cloud. That announcement restored free compute: $5 for new users’ first 30 days, followed after $5 or 30 days by $1 of non-rollover credits each month. Railway also reported serving more than 12 million deploys a month.[6]

Support was rebuilt around Central Station, an in-house combination of forum, helpdesk, roadmap and bounty payout system with bidirectional Slack and Discord integrations. Cooper described clustering customer feedback and support requests to identify emerging incidents and route context to relevant engineers. Swyx separately recalled helping Angelo with Central Station in 2023; that recollection describes a contribution rather than ownership of the complete system.[6][1]

On March 4, 2025, Jake Runzer introduced Railpack in beta as the successor to Nixpacks. Railway’s use of Nix tied package versions to repository commits: updating a commit to obtain one newer package could also change other dependencies. Nixpacks also collected build and runtime dependencies in a large /nix/store layer, limiting image-size reductions and cache control. Runzer qualified these as problems with Railway’s use of Nix.[5]

Railpack moved the builder from Rust to Go to use BuildKit libraries. It analyzes source code, creates a JSON-serializable build plan, then constructs a BuildKit graph defining the required layers. Mise resolves package versions and handles most installation; successful builds can lock dependencies, and BuildKit secrets keep secret variables out of build logs and final images. Railway reported 38% smaller base Node images and 77% smaller base Python images compared with Nixpacks.[5]

Railway announced a $100 million Series B on January 22, 2026; its January 23 changelog explicitly described the financing as closed. The changelog reported that over 2 million developers had shipped on Railway and tens of thousands of companies ran production on it. Railway called its direction an “Intelligent Cloud,” with shipped pieces including an Agent Skill and CLI tools for deployment, logs, metrics and environments.[3][4]

The same changelog introduced Focused PR Environments in beta. Railway matched changed files against each service’s watch paths and root directory, deploying affected services and services that referenced them instead of redeploying the entire project. It also shipped CLI service-configuration editing, crediting contributor Milo123459’s pull request.[4]

In the interview published in May, Cooper described a 35-person team supporting 3 million users and adding about 100,000 users a week. He said the vast majority of workloads ran on Railway’s bare metal, with AWS, GCP and other clouds retained for bursting; workloads moved back to owned hardware as space and power became available. After an upstream provider could not supply sufficient compute quota, he recalled rebuilding the network overlay over a weekend to span five environments, including Oracle, AWS, Railway and GCP.[1]

Cooper reported approximately three-month payback for moving from rented cloud compute to metal and roughly 70% margins on metal. Railway also used hardware-secured debt priced at prime plus a spread. Its deployment pipeline ran on Temporal, covering pre-commit hooks, signaling and queuing. Internally, Cooper estimated that Railway’s coding-agent spending had reached about $200,000 for the month under discussion, while identifying deployment of the resulting code as a continuing bottleneck.[1]

The May 19–20, 2026 outage exposed a remaining dependency beneath the multi-cloud design. In their incident report, Chandrika Khanduri and Cody De Arkland described a GCP production-account suspension as incorrect, while acknowledging earlier violation notices. Metal and AWS workloads remained online, but their edge routing depended on a GCP-hosted control-plane API. When cached routes expired, those workloads also became unreachable; at peak impact, all Railway workloads across all regions were unreachable.[9]

Railway reported an approximately eight-hour impact window, from 22:20 UTC on May 19 to 06:14 UTC on May 20, when the incident moved to monitoring; final resolution was recorded at 07:58 UTC. It accepted responsibility for the architecture that allowed one provider action to cascade across the platform. Announced remediation included removing the routing dependency, extending high-availability database shards across AWS and Metal, and taking GCP services out of the data plane’s hot path.[9]

On July 17, 2026, Victor Ramirez and Mahmoud Abdelwahab announced feature flags on every Railway project, accessible through the dashboard, CLI, SDK and MCP. Flags let an application select code paths at runtime, separating deployment from release. Percentage rollouts keep a stable assignment for each application-supplied user key. The SDK evaluates reads locally against a background-refreshed ruleset; conflicting matching rules or no match return the configured default, while an unreadable flag uses the fallback supplied in code.[8]

Mahmoud Abdelwahab’s August 13 announcement of dev.new extended Railway into generating the application itself. A prompt creates a project and boots a full VM using Railway Sandboxes infrastructure. Railway’s own railway-agent runs inside it, editing files and executing commands. The /app workspace and session transcript persist; after idle shutdown, the same volume is reattached when the user returns.[7]

The agent receives a pre-authenticated Railway CLI, Playwright with Chromium, and a ship tool that uploads /app for Railpack to build and start. Railway demonstrated the workflow with a multiplayer first-person shooter reportedly built and deployed in about twenty minutes. The agent tested two browser sessions, scripted damage, death and respawn checks, then repeated playtesting against the deployed URL. Abdelwahab qualified the result as a game enjoyable for a few minutes that would need much more work to sustain interest. dev.new was being tested with users; bringing the entire loop into Railway’s dashboard remained a future goal.[7]